Patent Application
20200372301
The recent discovery of adversarial attacks has raised concerns in regard to commercial deployment of deep learning models. Adversarial attacks probe a model and find a minimum perturbation to an input image that causes maximum degradation of the inference of a deep learning model, while simultaneously maintaining the perceived image integrity of the input image.
Several studies have demonstrated that adversarial attacks can fool audio-to-text models by inserting hidden voice commands into imperceptible audio signals. Similarly, computer vision models can be tricked by contaminating images with perturbations that are imperceptible to humans.
Humans and deep learning models perceive images and render decisions distinctly from each other. Attackers can take advantage of these distinctions to make subtle perturbations to the input image that are unperceivable to a human, but destroy a deep learning model.
In dentistry, adversarial attacks can be used to create malicious examples that compromise the diagnostic integrity of automated dental image classification, landmark detection, or segmentation models.
Several studies have attempted to automate some components of dental diagnostics, but have not identified a method to defend against adversarial attacks.
In order that the advantages of the invention will be readily understood, a more particular description of the invention briefly described above will be rendered by reference to specific embodiments illustrated in the appended drawings. Understanding that these drawings depict only typical embodiments of the invention and are not therefore to be considered limiting of its scope, the invention will be described and explained with additional specificity and detail through use of the accompanying drawings, in which:
It will be readily understood that the components of the invention, as generally described and illustrated in the Figures herein, could be arranged and designed in a wide variety of different configurations. Thus, the following more detailed description of the embodiments of the invention, as represented in the Figures, is not intended to limit the scope of the invention, as claimed, but is merely representative of certain examples of presently contemplated embodiments in accordance with the invention. The presently described embodiments will be best understood by reference to the drawings, wherein like parts are designated by like numerals throughout.
Embodiments in accordance with the invention may be embodied as an apparatus, method, or computer program product. Accordingly, the invention may take the form of an entirely hardware embodiment, an entirely software embodiment (including firmware, resident software, micro-code, etc.), or an embodiment combining software and hardware aspects that may all generally be referred to herein as a “module” or “system.” Furthermore, the invention may take the form of a computer program product embodied in any tangible medium of expression having computer-usable program code embodied in the medium.
Any combination of one or more computer-usable or computer-readable media may be utilized. For example, a computer-readable medium may include one or more of a portable computer diskette, a hard disk, a random access memory (RAM) device, a read-only memory (ROM) device, an erasable programmable read-only memory (EPROM or Flash memory) device, a portable compact disc read-only memory (CDROM), an optical storage device, and a magnetic storage device. In selected embodiments, a computer-readable medium may comprise any non-transitory medium that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device.
Computer program code for carrying out operations of the invention may be written in any combination of one or more programming languages, including an object-oriented programming language such as Java, Smalltalk, C++, or the like and conventional procedural programming languages, such as the “C” programming language or similar programming languages, and may also use descriptive or markup languages such as HTML, XML, JSON, and the like. The program code may execute entirely on a computer system as a stand-alone software package, on a stand-alone hardware unit, partly on a remote computer spaced some distance from the computer, or entirely on a remote computer or server. In the latter scenario, the remote computer may be connected to the computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection may be made to an external computer (for example, through the Internet using an Internet Service Provider).
The invention is described below with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and/or block diagrams, and combinations of blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions or code. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions/acts specified in the flowchart and/or block diagram block or blocks.
These computer program instructions may also be stored in a non-transitory computer-readable medium that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable medium produce an article of manufacture including instruction means which implement the function/act specified in the flowchart and/or block diagram block or blocks.
The computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions/acts specified in the flowchart and/or block diagram block or blocks.
Referring to
The method 100 may further include receiving 104 patient demographic data, such as age, gender, underlying health conditions (diabetes, heart disease, cancer, etc.). The method 100 may further include receiving 106 a patient treatment history. This may include a digital representation of periodontal treatments the patient has received, such as cleanings, periodontal scaling, root planning, cary fillings, root canals, orthodontia, oral surgery, or other treatments or procedures performed on the teeth, gums, mouth, or jaw of the patient.
The method 100 may include pre-processing 108 the image received at step 102. Note that in some embodiments, the image received is correctly oriented, obtained using a desired imaging modality, and free of contamination or defects such that pre-processing is not performed. In other embodiments, some or all of re-orienting, removing contamination (e.g., noise), transforming to a different imaging modality, and correcting for other defects may be performed at step 108. In some embodiments, step 108 may correct for distortion due to foreshortening, elongation, metal artifacts, and image noise due to poor image acquisition from hardware, software, or patient setup.
Step 108 may further include classifying the image, such as classifying which portion of the patient's teeth and jaw is in the field of view of the image. For example, a full-mouth series (FMX) typically includes images classified as Premolar2, Molar3, Anterior1, Anterior2, Anterior3, Jaw Region, Maxilla, and Mandible. For each of these, the view may be classified as being the left side or right side of the patients face.
In the following description reference to an “image” shall be understood to interchangeably reference either the original image from step 102 or an image resulting from the pre-processing of step 108.
The method 100 may further include processing 110 the image to identify patient anatomy. Anatomy identified may be represented as a pixel mask identifying pixels of the image that correspond to the identified anatomy and labeled as corresponding to the identified anatomy. This may include identifying individual teeth. As known in the field of dentistry, each tooth is assigned a number. Accordingly, step 110 may include identifying teeth in the image and determining the number of each identified teeth. Step 110 may further include identifying other anatomical features for each identified tooth, such as its cementum-enamel junction (CEJ), boney points corresponding to periodontal disease around the tooth, gingival margin (GM), junctional epithelium (JE), or other features of the tooth that may be helpful in characterizing the health of the tooth and the gums and jaw around the tooth.
The method 100 may further include detecting 112 features present in the anatomy identified at step 110. This may include identifying caries, measuring clinical attachment level (CAL), measuring pocket depth (PD), or identifying other clinical conditions that may indicate the need for treatment. The identifying step may include generating a pixel mask defining pixels in the image corresponding to the detected feature. The method 100 may further include generating 114 a feature metric, i.e. a characterization of the feature. This may include performing a measurement based on the pixel mask from step 112. Step 114 may further take as inputs the image and anatomy identified from the image at step 110. For example, CAL or PD of teeth in an image may be measured, such as using the machine-learning approaches described below (see discussion of
The result of steps 108, 110, 112, and 114 is an image that may have been corrected, labels, e.g. pixel masks, indicating the location of anatomy and detected features and a measurement for each detected feature. This intermediate data may then be evaluated 116 with respect to a threshold. In particular, this may include an automated analysis of the detected and measured features with respect to thresholds. For example, CAL or PD measured using the machine-learning approaches described below may be compared to thresholds to see if treatment may be needed. Step 116 may also include evaluating some or all of the images, labels, detected features, and measurements for detected features a machine learning model to determine whether a diagnosis is appropriate (see
If the result of step 116 is affirmative, then the method 100 may include processing 118 the feature metric from step 114 according to a decision hierarchy. The decision hierarchy may further operate with respect to patient demographic data from step 104 and the patient treatment history from step 106. The result of the processing according to the decision hierarchy may be evaluated at step 120. If the result is affirmative, than an affirmative response may be output 122. An affirmative response may indicate that the a course of treatment corresponding to the decision hierarchy is determined to be appropriate. If the result of processing 118 the decision hierarchy is negative, then the course of treatment corresponding to the decision hierarchy is determined not to be appropriate. The evaluation according to the method 100 may be performed before the fact, i.e. to determine whether to perform the course of treatment. The method 100 may also be performed after the fact, i.e. to determine whether a course of treatment that was already performed was appropriate and therefore should be paid for by insurance.
The method 200 may include evaluating 202 whether the treatment, SRP, has previously been administered within a threshold time period prior to a reference time that is either (a) the time of performance of the method 200 and (b) the time that the treatment was actually performed, i.e. the treatment for which the appropriateness is to be determined according to the method 100 and the method 200. For example, this may include whether SRP was performed within 24 months of the reference time.
If not, the method 200 may include evaluating 204 whether the patient is above a minimum age, such as 25 years old. If the patient is above the minimum age, the method 200 may include evaluating 206 whether the number of pockets having a depth exceeding a minimum pocket depth exceeds a minimum pocket number. For example, where the method 200 is performed to determine whether SRP is/was appropriate for a quadrant (upper left, upper right, lower left, lower right) of the patient's jaw, step 206 may include evaluating whether there are at least four teeth in that quadrant that collectively include at least 8 sites, each site including a pocket of at least 5 mm. Where the method 200 is performed to determine whether SRP is/was appropriate for an area that is less than an entire quadrant, step 206 may include evaluating whether there are one to three teeth that include at least 8 sites, each site including a pocket of at least 5 mm.
If the result of step 206 is positive, then an affirmative result is output, i.e. the course of treatment is deemed appropriate. If the result of step 206 is positive, then an affirmative result is output 208, i.e. the course of treatment is deemed appropriate. If the result of step 206 is negative, then a negative result is output 210, i.e. the course of treatment is deemed not to be appropriate.
If either of (a) SRP was found 202 to have been performed less than the time window from the reference time or (b) the patient is found 204 to be below the minimum age, the method 200 may include evaluating 212 whether a periodontal chart has been completed for the patient within a second time window from the reference time, e.g. six months. If the result of step 212 is positive, then processing may continue at step 206. If the result of step 212 is negative, then processing may continue at step 210.
The decision hierarchy of the method 200 is just one example. Decision hierarchies for other treatments may be evaluated according to the method 100, such as gingiovectomy; osseous mucogingival surgery; free tissue grafts; flap reflection or resection and debridement (with or without osseous recontouring); keratinized/attached gingiva preservation; alveolar bone reshaping; bone grafting (with or without use of regenerative substrates); guided tissue regeneration; alveolar bone reshaping following any of the previously-mentioned procedures; and tissue wedge removal for performing debridement, flap adaptation, and/or pocket depth reduction. Examples of decision hierarchies for these treatments are illustrated in the U.S. Provisional Application Ser. No. 62/848,905.
As described below, machine learning models, such as a CNN, may be used to perform various tasks described above with respect to the method 100. Training of the CNN may be simplified by ensuring that the images used are in a standard orientation with respect to the anatomy represented in the images. When images are obtained in a clinical setting they are often mounted incorrectly by a human before being stored in a database. The illustrated system 300 may be used to determine the orientation of anatomy in an image such that they may be rotated to the standard orientation, if needed, prior to subsequent processing with another CNN or other machine learning model.
A training algorithm 302 takes as inputs training data entries that each include an image 304 according to any of the imaging modalities described herein and an orientation label 306 indicating the orientation of the image, e.g. 0 degrees, 90 degrees, 180 degrees, and 270 degrees. The orientation label 306 for an image may be assigned by a human observing the image and determining its orientation. For example, a licensed dentist may determine the label 306 for each image 304.
The training algorithm 302 may operate with respect to a loss function 308 and modify a machine learning model 310 in order to reduce the loss function 308 of the model 310. In this case, the loss function 308 may be a function that increases with a difference between the angle estimated by the model 310 for the orientation of an image 304 and the orientation label 306 of the image.
In the illustrated embodiment, the machine learning model 310 is a convolution neural network. For example, the machine learning model 310 may be an encoder-based densely-connected CNN with attention-gated skip connections and deep-supervision. In the illustrated embodiment, the CNN includes six multi-scale stages 312 followed by a fully connected layer 314, the output 316 of the fully connected layer 314 being an orientation prediction (e.g. 0 degrees, 90 degrees, 180 degrees, or 270 degrees).
In some embodiment, each multi-scale stage 312 may contain three 3×3 convolutional layers, which may be paired with batch-normalization and leaky rectified linear units (LeakyReLU). The first and last convolutional layers of each stage 312 may be concatenated via dense connections which help reduce redundancy within the CNN by propagating shallow information to deeper parts of the CNN.
Each multi-scale network stage 312 may be downscaled by a factor of two at the end of each multi-scale stage 312 by convolutional downsampling. The second and fourth multi-scale stages 312 may be passed through attention gates 318a, 318b before being concatenated with the last layer. For example, the gating signal of attention gate 318a that is applied to the second stage 312 may be derived from the output of the fourth stage 312. The gating signal of attention gate 318b that is applied to the fourth stage 312 may be derived from the output of the sixth stage 312. Not all regions of the image 304 are relevant for determining orientation, so the attention gates 318a, 318b may be used to selectively propagate semantically meaningful information to deeper parts of the CNN.
In some embodiments, the input image 304 to the CNN is a raw 64×64 pixel image and the output 316 of the network is a likelihood score for each possible orientation. The loss function 308 may be trained with categorical cross entropy which considers each orientation to be an orthogonal category. Adam optimization may be used during training which automatically estimates the lower order moments and helps estimate the step size which desensitizes the training routine to the initial learning rate.
In at least one possible embodiment, the images 304 are 3D images, such as a CT scan. Accordingly, the 3×3 convolutional kernels of the multi-scale networks with 3×3×3 convolutional kernels. The output 316 of the CNN may therefore map to four rotational configurations 0, 90, 180, and 270 along the superior-inferior axis as well as one orthogonal orientation in the superior-inferior direction.
Because machine learning models may be sensitive to training parameters and architecture, for all machine learning models described herein, including the machine learning model 310, a first set of training data entries may be used for hyperparameter testing and a second set of training data entries not included in the first set may be used to assess model performance prior to utilization.
The training algorithm 302 for this CNN and other CNNs and machine learning models described herein may be implemented using PYTORCH. Training of this CNN and other CNNs and machine learning models described herein may be performed using a GPU, such as NVIDIA's TESLA GPUs coupled with INTEL XEON CPUs. Other machine learning tools and computational platforms may also be used.
Generating inferences using this machine learning model 310 and other machine learning models described herein may be performed using the same type of GPU used for training or some other type of GPU or other type of computational platform. In other embodiment, inferences using this machine learning model 310 or other machine learning models described herein may be generated by placing the machine learning model on an AMAZON web services (AWS) GPU instance. During deployment, a server may instantiate the machine learning model and preload the model architecture and associated weights into GPU memory. A FLASK server may then load an image buffer from a database, convert the image into a matrix, such as a 32-bit matrix, and load it onto the GPU. The GPU matrix may then be passed through the machine learning model in the GPU instance to obtain an inference, which may then be stored in a database. Where the machine learning model transforms an image or pixel mask, the transformed image or pixel mask may be stored in an image array buffer after processing of the image using the machine learning model. This transformed image or pixel mask may then be stored in the database as well.
In the case of the machine learning model 310 of
In dentistry, an FMX is often taken to gain comprehensive imagery of oral anatomy. Standard views are categorized by the anatomic region sequence indicating the anatomic region being viewed such as jaw region, maxilla, or mandible and an anatomic region modifier sequence indicating a particular sub-region being viewed such as premolar 2, molar 3, anterior 1, anterior 2, and anterior 3. In addition, each anatomic region sequence and anatomic region sequence modifier has a laterality indicting which side of the patient is being visualized, such as left (L), right (R), or ambiguous (A). Correct identification, diagnosis, and treatment of oral anatomy and pathology rely on accurate pairing of FMX mounting information of each image.
In some embodiment, the system 400 may be used to train a machine learning model to estimate the view of an image. Accordingly, the output of the machine learning model for a given input image will be a view label indicating an anatomic region sequence, anatomic region sequence modifier, and laterality visualized by the image. In some embodiments, the CNN architecture may include an encoder-based residually connected CNN with attention-gated skip connections and deep-supervision as described below.
In the system 400, A training algorithm 402 takes as inputs training data entries that each include an image 404 according to any of the imaging modalities described herein and a view label 406 indicating which of the view the image corresponds to (anatomic region sequence, anatomic region sequence modifier, and laterality). The view label 406 for an image may be assigned by a human observing the image and determining which of the image views it is. For example, a licensed dentist may determine the label 406 for each image 404.
The training algorithm 402 may operate with respect to a loss function 408 and modify a machine learning model 410 in order to reduce the loss function 408 of the model 410. In this case, the loss function 408 may be a function that is zero when a view label output by the model 410 for an image 406 matches the view label 406 for that image 404 and is non-zero, e.g. 1, when the view label output does not match the view label 406. Inasmuch as there are three parts to each label (anatomic region sequence, anatomic region modifier sequence, and laterality) there may be three loss functions 408, one for each part that is zero when the estimate for that part is correct and non-zero, e.g. 1, when the estimate for that part is incorrect. Alternatively, the loss function 408 may output a single value decreases with the number of parts of the label that are correct and increase with the number of parts of the label that are incorrect
The training algorithm 402 may train a machine learning model 410 embodied as a CNN. In the illustrated embodiment, the CNN includes seven multi-scale stages 312 followed by a fully connected layer 414 that outputs an estimate for the anatomic region sequence, anatomic region modifier sequence, and laterality of an input image 404. Each multi-scale stage 412 may contain three 3×3 convolutional layers that may be paired with batchnormalization and leaky rectified linear units (LeakyReLU). The first and last convolutional layers of a stage 412 may be concatenated via residual connections which help reduce redundancy within the network by propagating shallow information to deeper parts of the network.
Each multi-scale stage 412 may be downscaled by a factor of two at the end of each multi-scale stage 412, such as by max pooling. The third and fifth multi-scale stages 412 may be passed through attention gates 418a, 418b, respectively, before being concatenated with the last stage 412. For example, the gating signal of attention gate 418a that is applied to the output of the third stage 412 may be derived from the fifth stage 412 and the gating signal applied by attention gate 418b to the output of the fifth stage 412 may be derived from the seventh stage 412. Not all regions of the image are relevant for classification, so attention gates 418a, 418b may be used to selectively propagate semantically meaningful information to deeper parts of the network.
The input images 404 may be raw 128×128 images, which may be rotated to a standard orientation according to the approach of
In at least one possible embodiment, the images 404 are 3D images, such as a CT scan. Accordingly, the 3×3 convolutional kernels of the multi-scale stages 412 may be replaced with 3×3×3 convolutional kernels. The output of the machine learning model 4120 in such embodiments may be a mapping of the CT scan to one of a number of regions within the oral cavity, such as the upper right quadrant, upper left quadrant, lower left quadrant, and lower right quadrant.
The training algorithm 402 and utilization of the trained machine learning model 410 may be implemented using PYTORCH and AWS GPU instances in the same manner as described above with respect to
In some embodiment, the system 500 may be used to train a machine learning model to output an improved quality image for a given input image. In order to establish the correct diagnosis from dental images, it is often useful to have high resolution, high contrast, and artifact free images. It can be difficult to properly delineate dental anatomy if image degradation has occurred due to improper image acquisition, faulty hardware, patient setup error, or inadequate software. Poor image quality can take many forms such as noise contamination, poor contrast, or low resolution. The illustrated system 500 may be used to solve this problem.
In the system 500, A training algorithm 502 takes as inputs contaminated images 504 and real images 506. As for other embodiments, the images 504, 506 may be according to any of the imaging modalities described herein. The images 504 and 506 are unpaired in some embodiments, meaning the real images 506 are not uncontaminated versions of the contaminated images 504. Instead, the real images 506 may be selected from a repository of images and used to assess the realism of synthetic images generated using the system 500. The contaminated images 504 may be obtained by adding contamination to real images in the form of noise, distortion, or other defects. The training algorithm 502 may operate with respect to one or more loss functions 508 and modify a machine learning model 510 in order to reduce the loss functions 508 of the model 510.
In the illustrated embodiment, the machine learning model 510 may be embodied as a generative adversarial network (GAN) including a generator 512 and a discriminator 514. The generator 512 may be embodied as an encoder-decoder generator including seven multi-scale stages 516 in the encoder and seven multi-scale stages 518 in the decoder (the last stage 516 of the encoder being the first stage of the decoder). The discriminator 514 may include five multi-scale stages 522.
Each multi-scale stage 516, 518 within the generator 512 may use 4×4 convolutions paired with batchnormalization and rectified linear unit (ReLU) activations. Convolutional downsampling may be used to downsample each multi-scale stage 516 and transpose convolutions may be used between the multi-scale stages 518 to incrementally restore the original resolution of the input signal. The resulting high-resolution output channels of the generator 512 may be passed through a 1×1 convolutional layer and hyperbolic tangent activation function to produce a synthetic image 520. At each iteration, the synthetic image 520 and a real image 506 from a repository of images may be passed through the discriminator 514.
The discriminator 514 produces as an output 524 a realism matrix that is an attempt to differentiate between real and fake images. The realism matrix is a matrix of values, each value being an estimate as to which of the two input images is real. The loss function 508 may then operate on an aggregation of the values in the realism matrix, e.g. average of the values, a most frequently occurring value of the values, or some other function. The closer the aggregation is to the correct conclusion (determining that the synthetic image 520 is fake), the lower the output of the loss function 508. The realism matrix may be preferred over a conventional single output signal discriminator because it is better suited to capture local image style characteristics and it is easier to train.
In some embodiments, the loss functions 508 utilize level 1 (L1) loss to help maintain the spatial congruence of the synthetic image 520 and real image 506 and adversarial loss to encourage realism. The generator 512 and discriminator 514 may be trained simultaneously until the discriminator 514 can no longer differentiate between synthetic and real images or a Nash equilibrium has been reached.
In at least one possible embodiment, the system 500 may operate on three-dimensional images 504, 506, such as a CT scan. This may include replacing the 4×4 convolutional kernels with 4×4×4 convolutional kernels and replacing the 1×1 convolutional kernels with 1×1×1 convolutional kernels.
The training algorithm 502 and utilization of the trained machine learning model 510 may be implemented using PYTORCH and AWS GPU instances in the same manner as described above with respect to
The system 600 may be used to train a machine learning model 610, e.g. a cyclic GAN, to transform an image obtained using one image modality to an image from another image modality. Examples of transforming between two-dimensional imaging modalities may include transforming between any two of the following: an X-ray, CBCT image, a slice of a CT scan, an intra-oral photograph, cephalometric, panoramic, or other two-dimensional imaging modality. In some embodiments, the machine learning model 610 may transform between any two of the following three-dimensional imaging modalities, such as a CT scan, magnetic resonance imaging (MM) image, a three-dimensional optical image, LIDAR (light detection and ranging) point cloud, or other three-dimensional imaging modality. In some embodiments, the machine learning model 610 may be trained to transform between any one of the two-dimensional imaging modalities and any one of the three-dimensional imaging modalities. In some embodiments, the machine learning model 610 may be trained to transform between any one of the three-dimensional imaging modalities and any one of the two-dimensional imaging modalities.
In some embodiments, the machine learning model 610 may be trained to translate between a first imaging modality that is subject to distortion (e.g., foreshortening or other type of optical distortion and a second imaging modality that is less subject to distortion. Deciphering dental pathologies on an image may be facilitated by establishing absolute measurements between anatomical landmarks (e.g., in a standard units of measurement, such as mm). Two-dimensional dental images interpret a three-dimensional space by estimating x-ray attenuation along a path from the target of an x-ray source to a photosensitive area of film or detector array. The relative size and corresponding lengths of any intercepting anatomy will be skewed as a function of their position relative to the x-ray source and imager. Furthermore, intra-oral optical dental images capture visual content by passively allowing scattered light to intercept a photosensitive detector array. Objects located further away from the detector array will appear smaller than closer objects, which makes estimating absolute distances difficult. Correcting for spatial distortion and image contamination can make deciphering dental pathologies and anatomy on x-ray, optical, or CBCT images more accurate. The machine learning model 610 may therefore be trained to translate between a distorted source domain and an undistorted target domain using unpaired dental images.
The transformation using the machine learning model 610 may be performed on an image that has been reoriented using the approach of
In the system 600, A training algorithm 602 takes as inputs images 604 from a source domain (first imaging modality, e.g., a distorted image domain) and images 606 from a target domain (second imaging modality, e.g., a non-distorted image domain or domain that is less distorted than the first domain). The images 604 and 606 are unpaired in some embodiments, meaning the images 606 are not transformed versions of the images 504 or paired such that an image 604 has a corresponding image 606 visualizing the same patient's anatomy. Instead, the images 506 may be selected from a repository of images and used to assess the transformation of the images 604 using the machine learning model 610. The training algorithm 502 may operate with respect to one or more loss functions 608 and modify a machine learning model 610 in order to reduce the loss functions 608 of the model 610.
Training of the machine learning model 610 may be performed by the training algorithm 602 as follows:
(Step 1) An image 604 in the source domain is input to generator 612 to obtain a synthetic image 622 in the target domain.
(Step 2) The synthetic image 622 and an unpaired image 606 from the target domain are input to the discriminator 614, which produces a realism matrix output 616 that is the discriminator's estimate as to which of the images 622, 606 is real.
(Step 3) Loss functions LF1 and LF2 are evaluated. Loss function LF1 is low when the output 616 indicates that the synthetic image 622 is real and that the target domain image 606 is fake. Since the output 616 is a matrix, the loss function LF1 may be a function of the multiple values (average, most frequently occurring value, etc.). Loss function LF2 is low when the output 616 indicates that the synthetic image 622 is fake and that the target domain image 606 is real. Thus, the generator 612 is trained to “fool” the discriminator 614 and the discriminator 614 is trained to detect fake images. The generator 612 and discriminator 614 may be trained concurrently.
(Step 4) The synthetic image 622 is input to the generator 618. The generator 618 transforms the synthetic image 622 into a synthetic source domain image 624.
(Step 5) A loss function LF3 is evaluated according to a comparison of the synthetic source domain image 624 and the source domain image 604 that was input to the generator 612 at Step 1. The loss function LF3 decreases with similarity of the images 604, 622.
(Step 6) A real target domain image 606 (which may be the same as or different from that input to the discriminator 614 at Step 2, is input to the generator 618 to obtain another synthetic source domain image 624. This synthetic source domain image 624 is input to the discriminator 620 along with a source domain image 604, which may be the same as or different from the source domain image 604 input to the generator 612 at Step 1.
(Step 7) The output 626 of the discriminator 620, which may be a realism matrix, is evaluated with respect to a loss function LF4 and a loss function LF5. Loss function LF4 is low when the output 626 indicates that the synthetic image 624 is real and that the source domain image 604 is fake. Since the output 626 is a matrix, the loss function LF4 may be a function of the multiple values (average, most frequently occurring value, etc.). Loss function LF5 is low when the output 626 indicates that the synthetic image 624 is fake and that the source domain image 604 is real.
(Step 8) The synthetic image 624 obtained at Step 6 is input to the generator 612 to obtain another synthetic target domain image 622.
(Step 9) A loss function LF6 is evaluated according to a comparison of the synthetic target domain image 622 from Step 8 and the target domain image 606 that was input to the generator 618 at Step 6. The loss function LF6 decreases with similarity of the images 606, 622.
(Step 10) Model parameters of the generators 612, 618 and the discriminators 614, 620 are tuned according to the outputs of the loss functions LF1, LF2, LF3, LF4, LF5, LF6, and LF7.
Steps 1 through 10 may be repeated until an ending condition is reached, such as when the discriminators 616, 620 can no longer distinguish between synthetic and real images (e.g., only correct 50 percent of the time), a Nash equilibrium is reached, or some other ending condition is reached.
Since the machine learning model 610 trains on un-paired images, a conventional L1 loss may be inadequate because the source and target domains are not spatially aligned. To promote spatial congruence between the source input image 604 and synthetic target image 622, the illustrated reverse GAN network (generator 618 and discriminator 620) may be used in combination with the illustrated forward GAN network (generator 612 and discriminator 614). Spatial congruence is therefore encouraged by evaluating L1 loss (loss function LF3) at Step 5 and evaluating L1 loss (loss function LF6) at Step 9.
Once training is ended, the generator 612 may be used to transform an input image in the source domain to obtain a transformed image in the target domain. The discriminators 616, 620 and the second generator 618 may be ignored or discarded during utilization.
The training algorithm 602 and utilization of the trained machine learning model 610 may be implemented using PYTORCH and AWS GPU instances in the same manner as described above with respect to
In at least one possible embodiment, the system 600 operates on three-dimensional images, such as a CT, by replacing two-dimensional convolutional kernels (e.g., 4×4 and 1×1) with three-dimensional convolution kernels (e.g., 4×4×4 or 1×1×1).
In the system 700, A training algorithm 702 takes as inputs training data entries that each include an image 704 and labels 706a for teeth represented in that image. For example, the labels 706a may be a tooth label mask in which pixel positions of the image 704 that correspond to a tooth are labeled as such, e.g. with the tooth number of a labeled tooth. The labels 706a for an image may be generated by a licensed dentist. The training algorithm 702 may further make use of unpaired labels 706b, i.e., pixels masks for images of real teeth, such as might be generated by a licensed dentist that do not correspond to the images 704 or labels 706a.
The training algorithm 702 may operate with respect to one or more loss functions 708 and modify a machine learning model 710 in order to train the machine learning model 710 to label teeth in a given input image. The labeling performed using the machine learning model 710 may be performed on an image that has been reoriented using the approach of
In the illustrated embodiment, the machine learning model 710 includes a GAN including a generator 712 and a discriminator 714. The discriminator 714 may have an output 716 embodied as a realism matrix that may be implemented as for other realism matrices in other embodiments as described above. The output of the generator 712 may also be input to a classifier 718 trained to produce an output 720 embodied as a tooth label, e.g. pixel mask labeling a portion of an input image estimated to include a tooth.
As for other GAN disclosed herein, the generator 712 may include seven multi-scale stage deep encoder-decoder generator, such as using the approach described above with respect to the generator 512. For the machine learning model 710, the output channels of the generator 712 may be passed through a 1×1 convolutional layer as for the generator 512. However, the 1×1 convolution layer may further include a sigmoidal activation function to produce tooth labels. The generator 712 may likewise have stages of a different size than the generator 512, e.g., an input stage of 256×256 with downsampling by a factor of two between stages.
The discriminator 714 may be implemented using the approach described above for the discriminator 514. However, in the illustrated embodiment, the discriminator 514 includes four layers, though five layers as for the discriminator 514 may also be used.
The classifier 718 may be embodied as an encoder including six multi-scale stages 722 coupled to a fully connected layer 724, the output 720 of the fully connected layer 314 being a tooth label mask. In some embodiments, each multi-scale stage 722 may contain three 3×3 convolutional layers, which may be paired with batch-normalization and leaky rectified linear units (LeakyReLU). The first and last convolutional layers of each stage 722 may be concatenated via dense connections which help reduce redundancy within the CNN by propagating shallow information to deeper parts of the CNN. Each multi-scale network stage 722 may be downscaled by a factor of two at the end of each multi-scale stage 722 by convolutional downsampling.
Training of the machine learning model 710 may be performed by the training algorithm 702 according to the following method:
(Step 1) An image 704 is input to the generator 712, which outputs synthetic labels 726 for the teeth in the image 704. The synthetic labels 726 and unpaired tooth labels 706b from a repository are input to the discriminator 714. The discriminator 714 outputs a realism matrix with each value in the matrix being an estimate as to which of the input labels 726, 706b is real.
(Step 2) Input data 728 is input to the classifier 718, the input data 728 including layers including the original image 704 concatenated with the synthetic label 726 from Step 1. In response, the classifier 718 outputs its own synthetic label on its output 720.
(Step 3) The loss functions 708 are evaluated. This may include a loss function LF1 based on the realism matrix output at Step 1 such that the output of LF1 decreases with increase in the number of values of the realism matrix that indicate that the synthetic labels 726 are real. Step 3 may also include evaluating a loss function LF2 based on the realism matrix such that the output of LF2 decreases with increase in the number of values of the realism matrix that indicate that the synthetic labels 726 are fake. Step 3 may include evaluating a loss function LF3 based on a comparison of the synthetic label output by the classifier 718 and the tooth label 706a paired with the image 704 processed at Step 1. In particular, the output of the loss function LF3 may decrease with increasing similarity of the synthetic label output from the classifier 718 and the tooth label 706a.
(Step 4) The training algorithm 702 may use the output of loss function LF1 to tune parameters of the generator 712, the output of loss function LF2 to tune parameters of the discriminator 714, and the output of the loss function LF3 to tune parameters of the classifier 718. In some embodiments, the loss functions 708 are implemented as an objective function that utilizes a combination of softdice loss between the synthetic tooth label 726 and the paired truth tooth label 706a, adversarial loss from the discriminator 714, and categorical cross entropy loss from the classifier 718.
Steps 1 through 4 may be repeated such that the generator 712, discriminator 714, and classifier 718 are trained simultaneously. Steps 1 through 4 may continue to be repeated until an end condition is reached, such as until loss function LF3 meets a minimum value or other ending condition and LF2 is such that the discriminator 714 identifies the synthetic labels 726 as real 50 percent of the time or Nash equilibrium is reached.
During utilization, the discriminator 716 may be ignored or discarded. Images may then be processed by the generator 712 to obtain a synthetic label 726, which is then concatenated with the image to obtain data 728, which is then processed by the classifier 718 to obtain one or more tooth labels.
The training algorithm 702 and utilization of the trained machine learning model 710 may be implemented using PYTORCH and AWS GPU instances in the same manner as described above with respect to
In at least one possible embodiment, the system 700 operates on three-dimensional images, such as a CT, by replacing two-dimensional convolutional kernels (e.g., 4×4 and 1×1) with three-dimensional convolution kernels (e.g., 4×4×4 or 1×1×1).
In the system 800, A training algorithm 802 takes as inputs training data entries that each include an image 804a and labels 804b for teeth represented in that image, e.g., pixel masks indicating portions of the image 804a corresponding to teeth. The labels 804b for an image 804a may be generated by a licensed dentist or automatically generated using the tooth labeling system 700 of
As described below, two versions of the feature label 806 may be used. An non-dilated version is used in which only pixels identified as corresponding to the anatomical feature of interest are labeled. A dilated version is also used in which the pixels identified as corresponding to the anatomical feature of interest are dilated: a mask is generated that includes a probability distribution for each pixel rather than binary labels. Pixels that were labeled in the non-dilated version will have the highest probability values, but adjacent pixels will have probability values that decay with distance from the labeled pixels. The rate of decay may be according to a gaussian function or other distribution function. Dilation facilitates training of a machine learning model 810 since a loss function 808 will increase gradually with distance of inferred pixel locations from labeled pixel locations rather than being zero at the labeled pixel locations and the same non-zero value at every other pixel location.
The training algorithm 802 may operate with respect to one or more loss functions 808 and modify a machine learning model 810 in order to train the machine learning model 810 to label the anatomical feature of interest in a given input image. The labeling performed using the machine learning model 810 may be performed on an image that has been reoriented using the approach of
In the illustrated embodiment, the machine learning model 810 includes a GAN including a generator 812 and a discriminator 814. The discriminator 814 may have an output 816 embodied as a realism matrix that may be implemented as for other realism matrices in other embodiments as described above. The output of the generator 812 may also be input to a classifier 818 trained to produce an output 820 embodied as a label of the anatomical feature of interest, e.g. pixel mask labeling a portion of an input image estimated to correspond to the anatomical feature of interest. The generator 812 and discriminator 814 may be implemented according to the approach described above for the generator 712 and discriminator 714. The classifier 818 may be implemented according to the approach described above for the classifier 718.
Training of the machine learning model 810 may be performed by the training algorithm 802 as follows:
(Step 1). The image 804a and tooth label 804b are concatenated and input to the generator 812. Concatenation in this and other systems disclosed herein may include inputting two images (e.g., the image 804a and tooth label 804b) as different layers to the generator 812, such as in the same manner that different color values (red, green, blue) of a color image may be processed by a CNN according to any approach known in the art. The generator 812 may output synthetic labels 822 (e.g., pixel mask) of the anatomical feature of interest based on the image 804a and tooth label 804b.
(Step 2) The synthetic labels 822 and real labels 824 (e.g., an individual pixel mask from a repository including one or more labels) are then input to the discriminator 814. The real labels 824 are obtained by labeling the anatomical feature of interest in an image that is not paired with the image 804a from Step 1. The discriminator 814 produces a realism matrix at its output 816 with each value of the matrix indicating whether the synthetic label 822 is real or fake. In some embodiments, the real labels 824 may be real labels that have been dilated using the same approach used to dilate the feature labels 806 to obtain the dilated feature labels 806. In this manner, the generator 812 may be trained to generate dilated synthetic labels 822.
(Step 3) The image 804a, tooth label 804b, and synthetic labels 822 are concatenated to obtain a concatenated input 826, which is then input to the classifier 818. The classifier 818 processes the concatenated input 826 and produces output labels 828 (pixel mask) that is an estimate of the pixels in the image 804a that correspond to the anatomical feature of interest.
(Step 4) The loss functions 808 are evaluated with respect to the outputs of the generator 812, discriminator 814, and classifier 818. This may include evaluating a loss function LF1 based on the realism matrix output by the discriminator 814 at Step 2 such that the output of LF1 decreases with increase in the number of values of the realism matrix that indicate that the synthetic labels 822 are real. Step 4 may also include evaluating a loss function LF2 based on the realism matrix such that the output of LF2 decreases with increase in the number of values of the realism matrix that indicate that the synthetic labels 822 are fake. Step 4 may include evaluating a loss function LF3 based on a comparison of the synthetic label 822 output by the generator 812 and the dilated tooth feature label 806. In particular, the output of the loss function LF3 may decrease with increasing similarity of the synthetic label 822 and the dilated tooth label 804b. Step 4 may include evaluating a loss function LF4 based on a comparison of the synthetic labels 828 to the non-dilated tooth label 804b such that the output of the loss function LF4 decreases with increasing similarity of the synthetic labels 828 and the non-dilated tooth label 804b.
(Step 5) The training algorithm 802 may use the output of loss function LF1 and LF3 to tune parameters of the generator 812. In particular, the generator 812 may be tuned to both generate realistic labels according to LF1 and to generate a probability distribution of a dilated tooth label according to LF3. The training algorithm 802 may use the output of loss function LF2 to tune parameters of the discriminator 814 and the output of the loss function LF4 to tune parameters of the classifier 818.
Steps 1 through 5 may be repeated such that the generator 812, discriminator 814, and classifier 818 are trained simultaneously. Steps 1 through 5 may continue to be repeated until an end condition is reached, such as until loss functions LF1, LF3, and LF4 meet a minimum value or other ending condition, which may include the discriminator 714 identifying the synthetic label 822 as real 50 percent of the time or Nash equilibrium is reached.
The training algorithm 802 and utilization of the trained machine learning model 810 may be implemented using PYTORCH and AWS GPU instances in the same manner as described above with respect to
In at least one possible embodiment, the system 800 operates on three-dimensional images, such as a CT, by replacing two-dimensional convolutional kernels (e.g., 4×4 and 1×1) with three-dimensional convolution kernels (e.g., 4×4×4 or 1×1×1).
During utilization to identify the anatomical feature of interest, the discriminator 814 may be ignored or discarded. Input images 804a with tooth labels 804b but without feature labels 806 are processed using the discriminator to obtain a synthetic labels 822. The image 804a, tooth labels 804b, and synthetic labels 822 are concatenated and input to the classifier 818 that outputs a label 828 that is an estimate of the pixels corresponding to the anatomical feature of interest.
Below are example applications of the system 800 to label anatomical features:
In the system 900, A training algorithm 802 takes as inputs training data entries that each include an image 904a and labels 904b, e.g., pixel masks indicating portions of the image 904a corresponding to teeth, CEJ, JE, B, or other anatomical features. The labels 904b for an image 904a may be generated by a licensed dentist or automatically generated using the tooth labeling system 700 of
Each training data entry may further include a CAL label 906 that may be embodied as a numerical value indicating the CAL for a tooth, or each tooth of a plurality of teeth, represented in the image. The CAL label 906 may be assigned to the tooth or teeth of the image by a licensed dentist.
The training algorithm 902 may operate with respect to one or more loss functions 908 and modify a machine learning model 910 in order to train the machine learning model 910 to determine one or more CAL values for one or more teeth represented in an input image.
In the illustrated embodiment, the machine learning model 910 is a CNN including seven multi-scale stages 912 followed by a fully connected layer 914 that outputs a CAL estimate 916, such as a CAL estimate 916 for each tooth identified in the labels 904b. Each multi-scale stage 912 may contain three 3×3 convolutional layers, paired with batchnormalization and leaky rectified linear units (LeakyReLU). The first and last convolutional layers of each stage 912 may be concatenated via dense connections which help reduce redundancy within the network by propagating shallow information to deeper parts of the network. Each multi-scale stage 912 may be downscaled by a factor of two at the end of each multi-scale stage by convolutional downsampling with stride 2. The third and fifth multi-scale stages 912 may be passed through attention gates 918a, 918b before being concatenated with the last multi-scale stage 912. The attention gate 918a applied to the third stage 912 may be gated by a gating signal derived from the fifth stage 912. The attention gate 918b applied to the fifth stage 912 may be gated by a gating signal derived from the seventh stage 912. Not all regions of the image are relevant for estimating CAL, so attention gates 918a, 918b may be used to selectively propagate semantically meaningful information to deeper parts of the network. Adam optimization may be used during training which automatically estimates the lower order moments and helps estimate the step size which desensitizes the training routine to the initial learning rate.
A training cycle of the training algorithm 902 may include concatenating the image 904a with the labels 904b of a training data entry and processing the concatenated data with the machine learning model 910 to obtain a CAL estimate 916. The CAL estimate 916 is compared to the CAL label 906 using the loss function 908 to obtain an output, such that the output of the loss function decreases with increasing similarity between the CAL estimate 916 and the CAL label 906. The training algorithm 902 may then adjust the parameters of the machine learning model 910 according to the output of the loss function 908. Training cycles may be repeated until an ending condition is reached, such as the loss function 908 reaching a minimum value or other ending condition being achieved.
The training algorithm 902 and utilization of the trained machine learning model 810 may be implemented using PYTORCH and AWS GPU instances in the same manner as described above with respect to
In at least one possible embodiment, the system 900 operates on three-dimensional images, such as a CT, by replacing two-dimensional convolutional kernels (e.g., 3×3 and 1×1) with three-dimensional convolution kernels (e.g., 3×3×3 or 1×1×1).
In the system 1000, a training algorithm 1002 takes as inputs training data entries that each include an image 1004a and labels 1004b, e.g., pixel masks indicating portions of the image 1004a corresponding to teeth, GM, CEJ, JE, B, or other anatomical features. The labels 1004b for an image 1004a may be generated by a licensed dentist or automatically generated using the tooth labeling system 700 of
The image 1004a may have been one or both of reoriented according to the approach of
The training algorithm 1002 may operate with respect to one or more loss functions 1008 and modify a machine learning model 1010 in order to train the machine learning model 1010 to determine one or more PD values for one or more teeth represented in an input image. In the illustrated embodiment, the machine learning model 1010 is a CNN that may be configured as described above with respect to the machine learning model 910.
A training cycle of the training algorithm 1002 may include concatenating the image 1004a with the labels 1004b of a training data entry and processing the concatenated data with the machine learning model 1010 to obtain a PD estimate 1016. The PD estimate 1016 is compared to the PD label 1006 using the loss function 1008 to obtain an output, such that the output of the loss function decreases with increasing similarity between the PD estimate 1016 and the PD label 1006. The training algorithm 1002 may then adjust the parameters of the machine learning model 1010 according to the output of the loss function 1008. Training cycles may be repeated until an ending condition is reached, such as the loss function 1008 reaching a minimum value or other ending condition being achieved.
The training algorithm 1002 and utilization of the trained machine learning model 1010 may be implemented using PYTORCH and AWS GPU instances in the same manner as described above with respect to
In at least one possible embodiment, the system 1000 operates on three-dimensional images, such as a CT, by replacing two-dimensional convolutional kernels (e.g., 3×3 and 1×1) with three-dimensional convolution kernels (e.g., 3×3×3 or 1×1×1).
In order to assess the extent of periodontal disease it is often useful to observe a multitude of dental images. Periodontal disease can be difficult to diagnosis on dental X-rays, CBCTs, and intra-oral images because periodontal disease relates to the cementum enamel junction (CEJ), junctional epithelium (JE), gingival margin (GM), boney point (B) on the maxilla or mandible, pocket depth (PD), gingival health, comorbidities, and clinical attachment level (CAL), which might not always be available. Furthermore, the contrast of soft tissue anatomy can be washed out from adjacent boney anatomy because bone attenuates more x-rays than soft tissue. Also, boney anatomy might not always be differentiated from other parts of the image or might be obfuscated by overlapping anatomy from adjacent teeth or improper patient setup and image acquisition geometry. To solve this problem, the illustrated system 1100 may be used in combination with the approaches of
For compatibility, all structured information (e.g. pixel mask labels, PD, and CAL values obtained using the approaches of
In the system 1100, A training algorithm 1102 takes as inputs training data entries that each include an image 1104a and labels 1104b, e.g., pixel masks indicating portions of the image 1104a corresponding to teeth, GM, CEJ, JE, B or other anatomical features. Each training data entry may further include a diagnosis 1106, i.e. a periodontal diagnosis that was determined by a licensed dentist to be appropriate for one or more teeth represented in the image 1104a.
The image 1104a may be an image that has been oriented according to the approach of
The labels 1104b for the image 1104a of a training data entry may be generated by a licensed dentist or automatically generated using the tooth labeling system 700 of
In some embodiments, other labels 1104b may be used. For example, a label 1104b may label a tooth in an image with a pixel mask indicating a past treatment with respect to that tooth. Other labels 1104b may indicate comorbidities of the patient represented in the image 1104a.
The training algorithm 1102 may operate with respect to one or more loss functions 1108 and modify a machine learning model 1110 in order to train the machine learning model 1110 to determine a predicted diagnosis for one or more teeth represented in an input image.
In the illustrated embodiment, the machine learning model 1110 includes nine multi-scale stages 1112 followed by a fully connected layer 1114 that outputs a predicted diagnosis 1116. Each multi-scale stage 1112 may contain three 3×3 convolutional layers, paired with batchnormalization and leaky rectified linear units (LeakyReLU). The first and last convolutional layers of each stage 1112 may be concatenated via dense connections which help reduce redundancy within the network by propagating shallow information to deeper parts of the network. Each multi-scale stage 1112 may be downscaled by a factor of two at the end of each multi-scale stage 1112, such as by convolutional downsampling with stride 2. The fifth and seventh multi-scale stages 1112 may be passed through attention gates 1118a, 1118b before being concatenated with the last stage 1112. The attention gate 1118a may be applied to the fifth stage 1112 according to a gating signal derived from the seventh stage 1112. The attention gate 1118b may be applied to the seventh stage 1112 according to a gating signal derived from the ninth stage 1112. Not all regions of the image are relevant for estimating periodontal diagnosis, so attention gates may be used to selectively propagate semantically meaningful information to deeper parts of the network. Adam optimization may be used during training which automatically estimates the lower order moments and helps estimate the step size which desensitizes the training routine to the initial learning rate.
A training cycle of the training algorithm 1102 may include concatenating the image 1104a with the labels 1104b of a training data entry and processing the concatenated data with the machine learning model 1110 to obtain a predicted diagnosis 1116. The predicted diagnosis is compared to the diagnosis 1106 using the loss function 1108 to obtain an output, such that the output of the loss function decreases with increasing similarity between the diagnosis 1116 and the diagnosis 1106, which may simply be a binary value (zero of correct, non-zero if not correct). The training algorithm 1102 may then adjust the parameters of the machine learning model 1110 according to the output of the loss function 1108. Training cycles may be repeated until an ending condition is reached, such as the loss function 1108 reaching a minimum value or other ending condition being achieved.
The training algorithm 1102 and utilization of the trained machine learning model 1110 may be implemented using PYTORCH and AWS GPU instances in the same manner as described above with respect to
In at least one possible embodiment, the system 1100 operates on three-dimensional images, such as a CT, by replacing two-dimensional convolutional kernels (e.g., 3×3 and 1×1) with three-dimensional convolution kernels (e.g., 3×3×3 or 1×1×1).
In another variation, several outputs from multiple image modalities or multiple images from a single modality are combined in an ensemble of networks to form a comprehensive periodontal diagnosis or treatment protocol. For example, a system 1100 may be implemented for each imaging modality of a plurality of imaging modalities. A plurality of images of the same patient anatomy according to the plurality of imaging modalities may then be labeled and processed according to their corresponding systems 1100. The diagnosis output for each imaging modality may then be unified to obtain a combined diagnosis, such as by boosting, bagging, or other conventional machine learning methods such as random forests, gradient boosting, or SVMs.
The system 1200 may be used to train a machine learning model to restore missing data to images for use in pre-processing an image at step 108 of the method 100. In some embodiment, missing data may be restored to an image using the approach of
In the system 1200, A training algorithm 1202 is trained using training data entries including an image 1204 and a randomly generated mask 1206 that defines portions of the image 1204 that are to be removed and which a machine learning model 1210 is to attempt to restore. As for other embodiments, the image 1204 of each training data entry may be according to any of the imaging modalities described herein. The training algorithm 1202 may operate with respect to one or more loss functions 1208 and modify the machine learning model 1210 in order to reduce the loss functions 1208 of the model 1210.
In the illustrated embodiment, the machine learning model 1210 is GAN including a generator 1212 and a discriminator 11214. The generator 1212 and discriminator may be implemented according to any of the approaches described above with respect to the generators 512, 612, 618, 712, 812 and discriminators 514, 614, 620, 714, 814 described above.
Training cycles of the machine learning model 1210 may include inputting the image 1204 and the random mask 1206 of a training data entry into the generator 1212. The mask 1206 may be a binary mask, with one pixel for each pixel in the image. The value of a pixel in the binary mask may be zero where that pixel is to be omitted from the image 1204 and a one where the pixel of the image 1204 is to be retained. The image as input to the generator 1212 may be a combination of the image 1204 and mask 1206, e.g. the image 1204 with the pixels indicated by the mask 1206 removed, i.e. replaced with random values or filled with a default color value.
The generator 1212 may be trained to output a reconstructed synthetic image 1216 that attempts to fill in the missing information in regions indicated by the mask 1206 with synthetic imaging content. In some embodiments, the generator 1212 learns to predict the missing anatomical information based on the displayed sparse anatomy in the input image 1204. To accomplish this the generator 1212 may utilize partial convolutions that only propagate information through the network that is near the missing information indicated by the mask 1206. In some embodiments, the binary mask 1206 of the missing information may be expanded at each convolutional layer of the network by one in all directions along all spatial dimensions.
In some embodiments, the generator 1212 is a six multi-scale stage deep encoder-decoder generator and the discriminator 124 is a five multi-scale level deep discriminator. Each convolutional layer within the encoder and decoder stage of the generator 1212 may uses 4×4 partial convolutions paired with batchnormalization and rectified linear unit (ReLU) activations. Convolutional downsampling may be used to downsample each multi-scale stage and transpose convolutions may be used to incrementally restore the original resolution of the input signal. The resulting high-resolution output channels may be passed through a 1×1 convolutional layer and hyperbolic tangent activation function to produce the synthetic reconstructed image 1216.
At each iteration, the synthetic image 1216 and a real image 1218 from a repository may be passed through the discriminator 1214, which outputs a realism matrix 1220 in which each value of the realism matrix 1220 is a value indicating which of the images 1216, 1218 is real.
The loss functions 1208 may be implementing using weighted L1 loss between the synthetic image 1216 and input image 1204 without masking. In some embodiments, the loss functions 1208 may further evaluate perceptual loss from the last three stages of the discriminator 1214, style loss based on the Gram matrix of the extracted features from the last three stages of the discriminator, and total variation loss. The discriminator 1214 may be pretrained in some embodiments such that it is not updated during training and only the generator 1212 is trained. In other embodiments, the generator 1212 and discriminator 1214 may be trained simultaneously until the discriminator 1214 can no longer differentiate between synthetic and real images or a Nash equilibrium has been reached.
During utilization, the discriminator 1214 may be discarded or ignored. An image to be reconstructed may be processed using the generator 1212. In some embodiments, a mask of the image may also be input as for the training phase. This mask may be generated by a human or automatically and may identify those portions of the image that are to be reconstructed. The output of the generator 1214 after this processing will be a synthetic image in which the missing portions have been filled in.
In some embodiments, multiple images from multiple image modalities or multiple images from a single modality may combined in an ensemble of networks to form a comprehensive synthetic reconstructed image. For example, each image may be processed using a generator 1214 (which may be trained using images of the imaging modality of the each image in the case of multiple imaging modalities) and the output of the generators 1214 may then be combined. The outputs may be combined by boosting, bagging, or other conventional machine learning methods such as random forests, gradient boosting, or state vector machines (SVMs).
In at least one possible embodiment, the system 1200 may operate on three-dimensional images 1204, such as a CT scan. This may include replacing the 4×4 convolutional kernels with 4×4×4 convolutional kernels and replacing the 1×1 convolutional kernels with 1×1×1 convolutional kernels.
The training algorithm 1202 and utilization of the trained machine learning model 1210 may be implemented using PYTORCH and AWS GPU instances in the same manner as described above with respect to
In at least one possible embodiment, the system 1200 operates on three-dimensional images, such as a CT, by replacing two-dimensional convolutional kernels (e.g., 4×4 and 1×1) with three-dimensional convolution kernels (e.g., 4×4×4 or 1×1×1).
Referring generally to
Referring to
In dentistry, adversarial attacks can be used to create malicious examples that compromise the diagnostic integrity of automated dental image classification, landmark detection, distortion correction, image transformation, text extraction, object detection, image denoising, or segmentation models. Additionally, images might be manually tampered with in photoshop or other image manipulation software to fool a clinician into incorrectly diagnosing disease
Adversarial attacks have highlighted cyber security threats to current deep learning models. Similarly, adversarial attacks on medical automation systems could have disastrous consequences to patient care. Because many industries are increasingly reliant on deep learning automation solutions, adversarial defense and detection systems have become a critical domain in the machine learning community.
There are two main types of adversarial defense approaches. One approach uses a screening algorithm to detect if an image is authentic and the other approach builds models that are robust against adversarial images. The quality of the defense system is dependent on the ability to create high quality adversarial examples.
To produce adversarial examples, attackers need to gain access to the system. Black box attacks assume no knowledge of model parameters or architecture. Grey box attacks have architectural information but have no knowledge of model parameters. White box attacks have a priori knowledge of model parameters and architecture. White box adversarial examples may be used to evaluate the defense of each model, since white box attacks are the most powerful.
For white box attacks, an adversarial attacking system may be implemented by building attacks directly on each victim model. In some embodiments, the attack system uses a novel variation of the projected gradient decent (PGD) method (Madry Kurakin), which is an iterative extension of the canonical fast gradient sign method (Goodfellow). PGD finds the optimal perturbation by performing a projected stochastic gradient descent on the negative loss function.
For grey box attacks, an adversarial attacking system may be implemented by building attacks on the output of each victim model. Since grey box attacks do not have access to the gradients of the model, the output of each victim model may be used to update the gradients of the attacking model. The attacking model therefore becomes progressively better at fooling the victim model through stochastic gradient decent.
For black box attacks, an adversarial attacking system may be implemented by building attacks on the output of many victim models. Since black box attacks do not have access to the gradients of any model, the output of many victim models are used to update the gradients of the attacking model. The attacking model therefore becomes progressively better at fooling the victim model through stochastic gradient decent.
The systems disclosed herein may use adaptation of a coevolving attack and defense mechanism. After each epoch in the training routine, new adversarial examples may be generated and inserted into the training set. The defense mechanism is therefore trained to be progressively better at accurate inference in the presence of adversarial perturbations and the attack system adapts to the improved defense of the updated model.
Referring specifically to
In the illustrated embodiment, the machine learning model 1310 is an adversarial detection CNN. The CNN may include attention-gated skip connections and deep-supervision. In the illustrated embodiment, the CNN includes nine multi-scale stages 1312 followed by a fully connected layer 1314 that outputs an authenticity score 1320. Each multi-scale stage 1312 may contain three 3×3 convolutional layers, paired with batchnormalization and leaky rectified linear units (LeakyReLU). The first and last convolutional layers of each stage 1312 may be concatenated via dense connections which help reduce redundancy within the network by propagating shallow information to deeper parts of the network. Each multi-scale stage 1312 may be downscaled by a factor of two at the end of each multi-scale stage 1312, such as by max pooling. The fifth and seventh multi-scale stages 1312 may be passed through attention gates 1318a, 1318b before being concatenated with the last (ninth) stage 1312. The attention gate 1318a may be applied to the fifth stage 1312 according to a gating signal derived from the seventh stage 1312. The attention gate 1318b may be applied to the seventh stage 1312 according to a gating signal derived from the ninth stage 1312. Not all regions of the image are relevant for estimating periodontal diagnosis, so attention gates may be used to selectively propagate semantically meaningful information to deeper parts of the network. Adam optimization may be used during training which automatically estimates the lower order moments and helps estimate the step size which desensitizes the training routine to the initial learning rate.
In some embodiments, the images 1304 input to the network may be embodied as a raw 512×512 image 1304 and the output of the network may be a likelihood score 1320 indicating a likelihood that the input image 1304 is an adversarial example. The loss function 1308 may therefore decrease with accuracy of the score. For example, where a high score indicates an adversarial input image, the loss function 1308 decreases with increase in the likelihood score 1320 when the input image 1304 is an adversarial image. The loss function 1308 would then increase with increase in the likelihood score 1320 when the input image 1304 is not an adversarial image. The loss function 1308 may be implemented with categorical cross entropy and Adam optimization may be used during training which automatically estimates the lower order moments and helps estimate the step size which desensitizes the training routine to the initial learning rate.
The adversarial images 1304 in the training data set may be generated with any of projected gradient decent image contamination, synthetically generated images, and manually manipulated images by licensed dentists. Because the adversarial detection machine learning model 1310 may be sensitive to training parameters and architecture, a validation set may be used for hyperparameter testing and a final hold out test set may be used to assess final model performance prior to deployment.
The training algorithm 1302 and utilization of the trained machine learning model 1310 may be implemented using PYTORCH and AWS GPU instances in the same manner as described above with respect to
In at least one possible embodiment, the system 1300 operates on three-dimensional images, such as a CT, by replacing two-dimensional convolutional kernels (e.g., 4×4 and 1×1) with three-dimensional convolution kernels (e.g., 4×4×4 or 1×1×1).
An adversarial network 1408 may receive an uncontaminated image 1410 and process the image 1410 to generate additive noise 1412 to contaminate the input image in order to deceive a victim machine learning model 1414. The victim model 1414 may be any machine learning model described herein or any machine learning model trained to transform images or generate inferences based on images. Each image 1410 may have an accurate prediction associated with an input image 1410 may be a prediction obtained by processing the input image 1410 using the victim model 1414 without added noise 1412 or according to labeling by some other means, such as by a human with expertise.
The noise 1412 is combined with the image 1410 to obtain the contaminated input image 1402 that is input to the detector 1404. The detector 1404 attempts to detect these adversarial images 1402 and discard them. Input images 1402 that are not found to be adversarial are then input to the machine learning model 1414 that outputs a prediction 1416. The prediction 1416 is more robust due to the presence of the detector 1404 inasmuch as there is more assurance that the image 1402 is not adversarial.
Referring to
The loss function of the adversarial network 1408 may be based on the prediction 1414, i.e. if the loss function decreases with increasing inaccuracy of the prediction. For example, the input image 1408 may be part of a training data entry including an accurate prediction. The difference between the prediction 1414 and the accurate prediction may therefore be evaluated to determine the output of the loss function that is used to update the adversarial network.
In some embodiments, the loss function is a loss function 1418 that has two goal criteria minimizing 1420 noise and minimizing 1422 model performance, i.e. maximizing inaccuracy of the prediction 1416. Accordingly, the loss function 1418 may be a function of inaccuracy of the prediction 1416 relative to an accurate prediction associated with the input image 1408 and is also be a function of the magnitude of the adversarial noise 1412. The loss function 1418 therefore penalizes the adversarial network 1408 according to the magnitude of the noise and rewards the adversarial network 1408 according to degradation of accuracy of the victim model 1414.
The adversarial network 1408 and its training algorithm may be implemented according to any of the machine learning models described herein. In particular, the adversarial network 1408 may be implemented as a generator according to any of the embodiments described herein. In some embodiments, the adversarial network 1408 utilizes a six multi-scale level deep encoder-decoder architecture. Each convolutional layer within the encoder and decoder stage of the networks may use three 3×3 convolutions paired with batchnormalization and rectified linear unit (ReLU) activations. Convolutional downsampling may be used to downsample each multi-scale level and transpose convolutions may be used to incrementally restore the original resolution of the input signal. The resulting high-resolution output channels may be passed through a 1×1 convolutional layer and hyperbolic tangent activation function to produce adversarial noise 1412, which may be in the form of an image, where each pixel is the noise to be added to the pixel at that position in the input image 1410. At each iteration, the adversarial noise 1412 may be added to an image 1410 from a repository of training data entries to obtain the contaminated input image 1402. The contaminated input image 1402 may then be processed using the victim model 1414. The training algorithm may update model parameters of the adversarial network 1408 according to the loss function 1418. In some embodiments, the loss function 1418 is a function of mean squared error (MSE) of the adversarial noise 1412 and inverse cross entropy loss of the victim prediction 1416 relative to an accurate prediction associated with the input image 1408. In some embodiments, the victim model 1414 (e.g., machine learning model 1310) and the adversarial network 1408 may be trained concurrently.
For example, a low amount of randomized noise may not be sufficient to interfere with the adversarial noise 1412, resulting in greater errors relative to an intermediate amount of noise that is greater than the low amount. Likewise, where a larger amount of noise greater than the intermediate amount is used, accuracy of the machine learning model 1414 may be degraded due to low image quality. Accordingly, the tuning algorithm may identify intermediate values for the noise parameters that balance adversarial noise disruption with image quality degradation.
In some embodiments, the modulator 1426 is a machine learning model. The machine learning model may be a generator, such as according to any of the embodiments for a generator described herein. The modulator 1426 may therefore be trained using a machine learning algorithm to generate noise suitable to disrupt the adversarial noise 1412. For example, training cycles may include generating a contaminated input image 1402 as described above, processing the contaminated input image 1402 using the modulator 1426 to obtain a modulated input. The modulated input is then processed using the model 1414 to obtain a prediction 1416. A loss function that decreases with increase in the accuracy of the prediction 1416 relative to the accurate prediction for the image 1410 used to generate the contaminated input image 1402 may then be used to tune the parameters of the modulator 1426.
In the illustrated embodiment, input images 1402, which may be contaminated with adversarial noise 1412 are processed using a dynamic machine learning model 1428. In this manner, the ability to train the adversarial network 1408 to deceive the model 1428 is reduced relative to a static machine learning model 1414.
The dynamic machine learning model 1428 may be implemented using various approaches such as:
Computing device 1500 may be used to perform various procedures, such as those discussed herein. Computing device 1500 can function as a server, a client, or any other computing entity. Computing device can execute one or more application programs, such as the training algorithms and utilization of machine learning models described herein. Computing device 1500 can be any of a wide variety of computing devices, such as a desktop computer, a notebook computer, a server computer, a handheld computer, tablet computer and the like.
Computing device 1500 includes one or more processor(s) 1502, one or more memory device(s) 1504, one or more interface(s) 1506, one or more mass storage device(s) 1508, one or more Input/Output (I/O) device(s) 1510, and a display device 1530 all of which are coupled to a bus 1512. Processor(s) 1502 include one or more processors or controllers that execute instructions stored in memory device(s) 1504 and/or mass storage device(s) 1508. Processor(s) 1502 may also include various types of computer-readable media, such as cache memory.
Memory device(s) 1504 include various computer-readable media, such as volatile memory (e.g., random access memory (RAM) 1514) and/or nonvolatile memory (e.g., read-only memory (ROM) 1516). Memory device(s) 1504 may also include rewritable ROM, such as Flash memory.
Mass storage device(s) 1508 include various computer readable media, such as magnetic tapes, magnetic disks, optical disks, solid-state memory (e.g., Flash memory), and so forth. As shown in
I/O device(s) 1510 include various devices that allow data and/or other information to be input to or retrieved from computing device 1500. Example I/O device(s) 1510 include cursor control devices, keyboards, keypads, microphones, monitors or other display devices, speakers, printers, network interface cards, modems, lenses, CCDs or other image capture devices, and the like.
Display device 1530 includes any type of device capable of displaying information to one or more users of computing device 1500. Examples of display device 1530 include a monitor, display terminal, video projection device, and the like.
A graphics-processing unit (GPU) 1532 may be coupled to the processor(s) 1502 and/or to the display device 1530, such as by the bus 1512. The GPU 1532 may be operable to perform convolutions to implement a CNN according to any of the embodiments disclosed herein. The GPU 1532 may include some or all of the functionality of a general-purpose processor, such as the processor(s) 1502.
Interface(s) 1506 include various interfaces that allow computing device 1500 to interact with other systems, devices, or computing environments. Example interface(s) 1506 include any number of different network interfaces 1520, such as interfaces to local area networks (LANs), wide area networks (WANs), wireless networks, and the Internet. Other interface(s) include user interface 1518 and peripheral device interface 1522. The interface(s) 1506 may also include one or more user interface elements 1518. The interface(s) 1506 may also include one or more peripheral interfaces such as interfaces for printers, pointing devices (mice, track pad, etc.), keyboards, and the like.
Bus 1512 allows processor(s) 1502, memory device(s) 1504, interface(s) 1506, mass storage device(s) 1508, and I/O device(s) 1510 to communicate with one another, as well as other devices or components coupled to bus 1512. Bus 1512 represents one or more of several types of bus structures, such as a system bus, PCI bus, IEEE 1594 bus, USB bus, and so forth.
For purposes of illustration, programs and other executable program components are shown herein as discrete blocks, although it is understood that such programs and components may reside at various times in different storage components of computing device 1500, and are executed by processor(s) 1502. Alternatively, the systems and procedures described herein can be implemented in hardware, or a combination of hardware, software, and/or firmware. For example, one or more application specific integrated circuits (ASICs) can be programmed to carry out one or more of the systems and procedures described herein.
This application claims the benefit of the following applications, all of which are hereby incorporated herein by reference: U.S. Provisional Application Ser. No. 62/848,905 filed May 16, 2019, and entitled SYSTEMS AND METHODS FOR PERIODONTAL DISEASE AUTOMATED DENTAL INSURANCE CLAIMS ADJUDICATION. U.S. Provisional Application Ser. No. 62/850,556 filed May 21, 2019, and entitled A BIG DATA PLATFORM FOR TRACKING DISPARATE DENTAL INFORMATION. U.S. Provisional Application Ser. No. 62/850,559 filed May 21, 2019, and entitled ADVERSARIAL DEFENSE PLATFORM FOR AUTOMATED DENTAL IMAGE CLASSIFICATION. U.S. Provisional Application Ser. No. 62/867,817 filed Jun. 27, 2019, and entitled SYSTEM AND METHODS FOR AUTOMATED CARIES CLASSIFICATION, SCORING, QUANTIFICATION, AND INSURANCE CLAIMS ADJUDICATION. U.S. Provisional Application Ser. No. 62/868,864 filed Jun. 29, 2019, and entitled SYSTEMS AND METHODS FOR ARTIFICIAL INTELLIGENCE-BASED DENTAL IMAGE TO TEXT GENERATION. U.S. Provisional Application Ser. No. 62/868,870 filed Jun. 29, 2019, and entitled AN AUTOMATED DENTAL PATIENT IDENTIFICATION PLATFORM. U.S. Provisional Application Ser. No. 62/916,966 filed Oct. 18, 2019, and entitled SYSTEMS AND METHODS FOR AUTOMATED ORTHODONTIC RISK ASSESSMENT, MEDICAL NECESSITY DETERMINATION, AND TREATMENT COURSE PREDICTION.
| Number | Date | Country | |
|---|---|---|---|
| 62850556 | May 2019 | US | |
| 62850559 | May 2019 | US | |
| 62867817 | Jun 2019 | US | |
| 62868864 | Jun 2019 | US | |
| 62868870 | Jun 2019 | US | |
| 62916966 | Oct 2019 | US |
