The present invention relates to an image processing technique in an ultrasonic diagnostic apparatus.
One of fetal diagnoses using the ultrasonic diagnostic apparatus is an examination in which a size of a part of a fetus is measured according to an ultrasonic image and a weight of the fetus is estimated by the following Expression 1.
EFW=1.07BPD3+3.00×10−1AC2×FL [Ex. 1]
where EFW is an estimated fetal weight (g), BPD is a biparietal diameter (cm), AC is abdominal circumference (cm), and FL is a femur length (cm).
In a measurement section image used for a fetal weight estimation, recommended conditions are indicated by Japan Ultrasonic Medical Association in Japan. In a measurement cross section of the biparietal diameter, which is one of objects to be measured, in the Journal of Medical Ultrasonics Vol. 30 No. 3 (2003) “Standardization of ultrasonic fetal measurement and Japanese reference values”, there is a disclosure of “a cross-section in which a midline echo of the fetal head is depicted in a center of the cross-section and a transparent septum pellucidum and a cisterna corpora quadrigemina are depicted”.
Since there is a possibility that target parts may be depicted with different sizes and an estimated fetal weight is miscalculated depending on acquired position and angle of the image of the head measurement section that satisfies the recommended conditions, it is important to accurately acquire a sectional image that satisfies the above features. Patent Literature 1 is available as a prior art for acquiring the measurement section image satisfying the above features without depending on an inspector. Patent Literature 1 discloses that “a luminance spatial distribution feature statistically characterizing a measurement reference image is learned in advance, and a sectional image having the nearest luminance spatial distribution characteristic among multiple sectional images acquired by a cross-section acquisition unit 107 is selected as the measurement reference image”.
Patent Literature 1: WO 2012/042808
In Patent Literature 1, in the actual measurement, there are restrictions on the position and angle at which the cross-sectional image is acquired by a posture of the fetus in an uterus, and a determination is made based on the whole luminance information on the acquired cross-sectional image. As a result, it is assumed that it is difficult to acquire the cross-sectional image that completely satisfies the features required for the measurement. In other words, there is no high possibility that the acquired image becomes a cross-sectional image most suitable for the measurement by a doctor.
In order to address the above problems, an object of the present invention is to provide an ultrasonic diagnostic apparatus and an image processing method, which are capable of extracting features to be satisfied as a measurement cross-section, classifying the extracted features according to importance, and displaying and selecting a cross-sectional image suitable for each measurement item.
In order to address the above problems, according to the present invention, there is provided an ultrasonic diagnostic apparatus including: an image processing unit that generates an acquired image of a tissue in a subject based on a signal acquired from a probe that transmits and receives ultrasonic waves; an input unit that accepts an instruction from a user; an adequacy determination unit that determines whether the acquired image is adequate as a measurement image used for measuring the subject included in the acquired image, or not; and an output unit that presents to the operator a result determined by the adequacy determination unit.
Further, in order to address the above object, according to the present invention, there is provided an image processing method for an ultrasonic diagnostic apparatus, in which the ultrasonic diagnostic apparatus generates an acquired image of a tissue in a subject based on a signal acquired from a probe that transmits and receives ultrasonic waves, determines whether the acquired image is adequate as a measurement image used for measuring the subject included in the acquired image, or not, and presents a determination result to an operator.
According to the present invention, the features to be satisfied as the measurement cross-section can be extracted, the extracted features can be classified according to importance, and the cross-sectional image suitable for each measurement item can be displayed and selected.
Embodiments of the present invention will be described below with reference to the drawings. In the following embodiments, a head measurement cross-section will be described as an example of a diagnostic target in an ultrasonic diagnostic apparatus, but the present embodiment is applicable to an abdomen measurement cross-section and a femur measurement cross-section in the same way.
As is apparent from the figure, septum pellucidums 2003, 2004, and cisterna corpora quadrigeminas 2005, 2006 are extracted on both sides of the midline 2002 in a head contour 2001.
A first embodiment is directed to an ultrasonic diagnostic apparatus configured to include an image processing unit that generates an acquired image of a tissue in a subject based on a signal acquired from a probe that transmits and receives ultrasonic waves, an input unit that accepts an instruction from a user, an adequacy determination unit that determines whether the acquired image is adequate as a measurement image used for measuring the subject included in the acquired image, or not, and an output unit that presents to the operator a result determined by the adequacy determination unit. In addition, the first embodiment is directed to an image processing method for the ultrasonic diagnostic apparatus, which generates an acquired image of a tissue in a subject based on a signal acquired from a probe that transmits and receives ultrasonic waves, determines whether the acquired image is adequate as a measurement image used for measuring the subject included in the acquired image, or not, and presents a determination result to an operator.
In the above configuration, when the user operates the probe 1001, the image processing unit 1005 receives the image data through the transmitting and receiving unit 1002, the analog to digital conversion unit 1003, and the beamforming processing unit 1004. The image processing unit 1005 generates a cross-sectional image as the acquired image, and the monitor 1006 displays the cross-sectional image. The image processing unit 1005, the adequacy determination unit 1007, and the control unit 1010 can be realized by a program executed by a central processing unit (CPU) 1011 which is a processing unit of a normal computer. Hereinafter, the adequacy determination unit 1007 and the presentation unit 1008 for presenting the result to the user will be described. The presentation unit 1008 can also be realized by a program of the CPU as with the adequacy determination unit 1007.
As described below, the adequacy determination unit 1007 extracts the first partial images with the predetermined shape and size from the acquired image, identifies the first partial image in which the measurement target part is depicted from the extracted first partial images, extracts second partial images with a predetermined shape and size from the first partial image in which the measurement target part is depicted, extracts the components included in the measurement target part from the extracted multiple second partial images, calculates an evaluation value as a result of checking the positional relationship of the extracted component with a reference value, calculates a mean luminance value for each of the components, and calculates the degree of adequacy indicating whether the acquired image is adequate as the measurement image, or not, with the use of the evaluation value of the component and the mean luminance value of each component.
Specifically, the measurement part detection unit 3002 and the component detection unit 3004 detect the measurement part and the component by template matching. Template images used for the template matching are created in advance from images to be used as the reference of the measurement cross-section and stored in an internal memory of the ultrasonic diagnostic apparatus, a storage unit of the computer, or the like.
As shown in
Each processing unit of the adequacy determination unit 1007 shown in
The measurement part detection unit 3002 detects the input image patch in which the measurement part is depicted by template matching from the input image patches extracted by the measurement part comparison region extraction unit 3001 and outputs the input image patch. In the case of detecting the head contour, the input image patches 5002 and 5003 are sequentially compared with the head contour template image 4006 to calculate the degree of similarity. The degree of similarity is defined as SSD (Sum of Squared Difference) shown in Expression 2 below.
In this example, I(x, y) is a luminance value at coordinates (x, y) of the input image patch, and T(x, y) is the luminance value at the coordinates (x, y) of the template image.
When the input image patch completely matches the head contour template image, the SSD becomes 0. An input image path having the smallest SSD is extracted among all of the input image patches, and output as a head contour extraction patch image. When there is no input image patch whose SSD value is equal to or less than a predetermined value, it is determined that the head contour is not depicted in the input image 5001, and the processing of the present embodiment is terminated. At this time, a fact that the measurement target part could not be detected may be presented to the user by a message or a mark on the monitor 1006, and the user may be urged to input another image.
The degree of similarity between the input image patch and the template image may be defined by SAD (Sum of Absolute Difference), NCC (Normalized Cross-Correlation), ZNCC (Zero-means Normalized Cross-Correlation) instead of SSD. In addition, the measurement part comparison region extraction unit 3001 generates a template image in which a rotation, an enlargement, and a reduction are combined together, thereby being capable of detecting the head contours depicted with various arrangements and sizes. Further, an edge extraction, a noise removal, and so on are applied to both of the template image and the input image patch as preprocessing, thereby being capable of improving a detection accuracy.
The component comparison region extraction unit 3003 further extracts multiple second partial images with a predetermined shape and size from the input image patch in which the measurement part detected by the measurement part detection unit 3002 is depicted, and outputs the multiple second partial images. In other words, as shown in
The component detection unit 3004 detects a measurement part image patch in which the component included in the measurement part is depicted by template matching from the measurement part image patches extracted by the component comparison region extraction unit 3003, and outputs the detected measurement part image patch. In the case of detecting the midline, the septum pellucidum, and the cisterna corpora quadrigemina which are located inside the head contours 4002 and 6001, similarly to the processing by the measurement part detection unit 3002, the component detection unit 3004 sequentially compares the measurement part image patch with the midline template image 4008, the septum pellucidum template image 4009, and the cisterna corpora quadrigemina template image 4010 to calculate the respective similarity degrees, and extracts the measurement part image patch having SSD of a predetermined value or lower.
Since the feature amount of the septum pellucidum template image 4009 and the cisterna corpora quadrigemina template image 4010 is larger than that of the midline template image 4008, it is desirable to detect the septum pellucidum template image 4009 and the cisterna corpora quadrigemina template image 4010 prior to the midline. As shown in
The placement recognition unit 3005 recognizes a positional relationship of the components identified by the component detection unit 3004. In the case of the head, as shown in
The luminance value calculation unit 3006 calculates a mean of the luminance values of pixels included in the components specified by the component detection unit 3004, and stores the mean in the component luminance table.
The adequacy calculation unit 3007 calculates the degree of adequacy as the measurement cross-section with reference to the component placement evaluation table 8001 and the component luminance table 9001, and outputs the calculated degree of adequacy. The degree of adequacy is represented by the following Expression 3.
In the Expression, E is the degree of adequacy, Pi is each evaluation value stored in the component placement evaluation table 8001, qj is each means luminance value stored in the component luminance table 9001, and ai and bi are weighting factors taking values between 0 and 1. E takes a value between 0 and 1.
Each weighting factor is stored in advance in an adequacy weighting factor table as shown in
The presentation unit 1008 presents the degree of adequacy calculated by the adequacy calculation unit 3007 to the user through the monitor 1006, and the process is completed.
In the ultrasonic diagnostic apparatus according to the present embodiment, the number of weeks of fetus designated by the user through the user input unit 1009 may be used as auxiliary information. Since how to depict a size of the measurement part, the luminance value, and so on differs depending on the number of weeks of fetus, an improvement in detection accuracy can be expected by using the template images of the same fetal week number in the measurement part detection unit 3002 and the component detection unit 3004. Further, the weighting factor of the adequacy weighting factor table 10001 is changed according to the number of weeks of fetus, thereby being capable of more appropriately calculating the degree of adequacy. The number of weeks of fetus may be designated by the user through the user input unit 1009, but the number of weeks of fetus estimated by using the results measured in different parts in advance may be used.
With the ultrasonic diagnostic apparatus according to the first embodiment described in detail above, the features to be satisfied as the measurement cross-section is classified according to importance, and the cross-sectional image that satisfies particularly the features high in importance can be selected.
The present embodiment is directed to an ultrasonic diagnostic apparatus capable of selecting an optimum image as a measurement cross-sectional image when multiple cross-sectional images are input. In other words, the present embodiment is directed to an ultrasonic diagnostic apparatus configured such that an image processing unit generates multiple cross-sectional images, an adequacy determination unit determines whether the multiple cross-sectional images are adequate, or not, and an output unit selects and presents a cross-sectional image determined to be most adequate by the adequacy determination unit. The configuration shown in
The adequacy determination unit 1007 performs each processing described in the first embodiment on each of the multiple cross-sectional images generated by the image processing unit 1005, and determines the degree of adequacy. The determination result is stored in an adequacy table as shown in
As a third embodiment, a description will be given of an embodiment of a configuration in which feature quantities of a measurement cross-section is identified by machine learning with a smaller processing amount, and determined whether to be adequate, or not. In other words, the present embodiment is directed to an ultrasonic diagnostic apparatus in which the adequacy determination unit includes a candidate partial image extraction unit that extracts a partial image with an arbitrary shape and size from the acquired image, a feature extractor that extracts a feature quantity included in the acquired image from the partial image, and a classifier that identifies and classifies the extracted feature quantity.
In the first embodiment, the components included in the measurement part and the measurement part are extracted by template matching, and the degree of adequacy is determined with the use of the positional relationship of the components and the means luminance value. However, very large throughput is needed for template matching of the multiple cross-sectional images. In the present embodiment, a description will be given of a convolution neural network which extracts and identifies feature quantities from an input image using a machine. The feature quantity may be identified by a Bayesian classification, a k-nearest neighbor algorithm, a support vector machine, or the like with the use of a predetermined index such as a luminance value, an edge, or a gradient. The convolutional neural network has been disclosed in detail, in LECUN et al, “Gradient-Based Learning Applied to Document Recognition,” in Proc. IEEE, vol 86, no 11, November, and so on.
In this expression, f is an activation function, and x is an output value of a two-dimensional filter.
Although Expression 4 is a sigmoid function, a rectified linear unit or Maxout may be used as the activation function. The purpose of the convolution layer is to obtain local features by blurring a part of the input image or emphasizing the edge. In the case of head measurement, as an example, W1 is set to 200 pixels, k is set to 5 pixels, and W2 is set to 196 pixels. In the next pooling layer, a maximum pooling shown in Expression 5 is applied to the feature map generated by the convolution layer to generate a pooling layer output 15003 of W3×W3 size.
in this example, P is a region of an s×s size extracted from the feature map at an arbitrary position, yi is a luminance value of each pixel included in the extracted region, and y′ is a luminance value of the pooling layer output.
In the case of the head measurement, s is set to 2 pixels as an example. As a pooling method, mean pooling or the like may be used. The feature map is reduced by the pooling layer, and robustness can be ensured against minute position changes of features in the image. The similar processing is performed also on the convolution layer and the pooling layer at a post stage to generate a pooling layer output 15005. The classifier 14003 is a neural network configured by a fully connected layer 15006 and an output layer 15007, and outputs a classification result as to whether the input image satisfies the feature as the measurement cross-section, or not. Units of each layer are completely connected to each other, and for example, one unit of the output layer and a unit of an intermediate layer at a preceding stage have a relationship expressed by the following Expression 6.
In this case, Oi is an output value of an ith unit of the output layer, g is an activation function, N is the number of units of the intermediate layer, Cij is a weighting factor between the jth unit of the intermediate layer and the ith unit of the output layer, rj is an output value of the jth unit of the intermediate layer, and d is a bias. cij and d are updated by a learning process to be described later so as to be able to discriminate whether the feature as the measurement cross-section is satisfied, or not.
Next, a process of causing the convolution neural network of
Next, a process of determining whether the cross-sectional image satisfies the feature as the measurement cross-section, or not, with the use of the convolution neural network which has completed the learning will be described. The candidate partial image extracting unit 14001 exhaustively extracts the partial images from the entire input cross-sectional image and outputs the partial images. As indicated by arrow lines in
As described above, according to the apparatus of the present embodiment, since an optimum cross-sectional image is selected as a measurement cross-sectional image from the multiple cross-sectional images, a trouble of the user repeatedly acquiring the image and confirming the calculation result of the degree of adequacy can be omitted.
The respective embodiments of the present invention have been described above. However, the present invention includes various modified examples. For example, in the above-mentioned embodiments, in order to easily understand the present invention, the specific configurations are described. However, the present invention does not always provide all of the configurations described above. For example, in the embodiments described above, the ultrasonic diagnostic apparatus including the probe or the like has been described as an example. However, the present invention can be applied to a signal processing device that executes processing subsequent to the image processing unit on the storage data of the storage device in which the obtained RF signal and so on are accumulated. Also, a part of one configuration example can be replaced with another configuration example, and the configuration of one embodiment can be added with the configuration of another embodiment. Also, in a part of the respective configuration examples, another configuration can be added, deleted, or replaced.
Furthermore, an example in which a program that realizes parts or all of the above-described respective configurations, functions, processors, controllers, and so on is created has been described. Alternatively, parts and all of those configurations may be realized by hardware, for example, designed by an integrated circuit.
Filing Document | Filing Date | Country | Kind |
---|---|---|---|
PCT/JP2015/066015 | 6/3/2015 | WO | 00 |