MASS SPICULES DETECTION AND TRACING FROM DIGITAL MAMMOGRAMS

Abstract

The present invention provides an algorithm to detect and trace the spicules of a mass density in digital mammograms using an adaptive threshold edge algorithm and a flood-fill segmentation algorithm. Elongation criteria are used to remove false edges that do not radiate from a central mass margin. The algorithm works on a central mass border and spicules feature map that contains a subset of the pixels from the source image, so processing time is fast enough for use in a mammography CAD server and for real-time computation within a digital mammography workstation.

Description

STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH OR DEVELOPMENT

Not Applicable.

REFERENCE TO SEQUENCE LISTING, A TABLE, OR A COMPUTER PROGRAM LISTING COMPACT DISC APPENDIX

Not Applicable.

BACKGROUND OF THE INVENTION

The present invention relates generally to the field of medical imaging systems. Particularly, the present invention relates to an algorithm for mass spicules detection, tracing and display from digital mammography images in conjunction with a mammography CAD (Computer-aided detection) server and a digital mammography workstation.

The U.S. patent Classification Definitions: 382/173 (class 382, Image Analysis, subclass 173 Image Segmentation); 378/37 (class 378, X-Ray or Gamma Ray System or Devices, subclass 37 Mammography).

Mass spicules are visible as lines radiating from the margin of a mass. Most breast carcinomas have a mammographic appearance with such a spiculated structure, so it is an important sign used by radiologists when making a cancer diagnosis. The quality of automatic spicules detection and tracing has a direct impact on the detection performance of a CAD system. Displaying this information on a CAD workstation helps radiologists understand the CAD markers, and is especially important to help the radiologist dismiss false positive markers, and to enforce a positive finding. Currently no CAD system or workstation displays computer generated spicules traces (see reference 1).

BRIEF SUMMARY OF THE INVENTION

The current invention provides an algorithm to detect and trace the spicules in digital mammograms using an adaptive threshold to obtain a spicules feature map which is followed by a flood-fill segmentation algorithm to obtain the spicules tracing. Elongation criteria are used to remove false edges that do not radiate from the central mass margin. The algorithm works on a central mass border and spicules feature map that contains a subset of the entire image, so processing time is fast, allowing use in a mammography CAD server, and real-time computation within a digital mammography workstation.

There are two main steps used to obtain the spicules from a mass candidate ROI (region of interest): (1) pre-process the ROI to obtain a central mass segmentation and spicules feature map; (2) region grow from the central mass border into the spicules feature map by a flood-fill segmentation algorithm to obtain individual spicules.

The algorithm can also be applied to breast ultrasound and breast MRI images to obtain stellar margin tracings.

BRIEF DESCRIPTION OF THE SEVERAL VIEWS OF THE DRAWING

FIG. 1 shows how the proposed method from this invention is used in a computer-aided detection and review system (see reference 2).

The detailed pre-processing steps are shown in FIG. 2.

FIG. 3 shows a flowchart of the operations used to obtain the spicules after pre-processing.

FIG. 4 shows some example results from the algorithm described in this invention.

DETAILED DESCRIPTION OF THE INVENTION

FIG. 1 shows how the proposed method from this invention is used in a computer-aided detection and workstation review system. A set of 4 or more mammogram images from a mammography study is inputted to such system. Preprocessing is performed to remove artifacts outside the breast skinline as border of the breast tissue. Initial mass candidates are selected as region of interests for further processing. The candidate selection uses the integrated intensity and morphologic algorithms in reference 2. The central mass density segmentation and spicules segmentation is performed at each mass candidate location. Based on the segmentation information, feature is extracted and the each candidate can be classified. The segmentation and classification results are displayed on a workstation review system.

FIG. 2 shows the detailed pre-processing steps to prepare for mass and its spicules segmentation. The original high resolution images and the breast skin line as masks are used in these pre-processing steps. The steps start from each mass candidate.

A ROI (region of interest) is extracted from the mass candidate. The size of the region is defined as 4 times of the estimated mass size obtained from initial candidate calculation. The ROI is then down-sampled (or up-sampled) to a fixed dimension (such as, 256×256).

If the ROI is partially on the breast skin-line or chest-wall edges, the pixels outside the breast are filled up by mirroring the pixels inside the breast.

The fixed dimension ROI is enhanced by the following steps: (1) band-pass filtering using Gaussian blur difference with kernel size 96 and 8; (2) produce a directional edge image from the ROI, for example, using a 4 or 8 directional wavelet filter; (3) add the band-pass image weighting by 4 and the directional edge image to form the enhanced ROI image; (4) modulate the image with a “hat” shape modular, such as, Gaussian weights to obtain the enhanced ROI.

Use flood-fill segmentation technique to the enhanced ROI to search for central mass segmentation in the range between the seed-pixel-value (center of the candidate ROI) and the maximum-pixel-value within the ROI. Then the candidate location is updated to the centroid of the mass segmentation.

Apply an adaptive threshold algorithm to the enhanced ROI and the edge image to obtain the spicules feature map. The output of the FIG. 2 is the mass segmentation and the spicules feature map.

FIG. 3 shows a flowchart of the operations used to obtain the spicules after pre-processing. The central mass segmentation and the spicules feature map are used in this algorithm as inputs.

The central mass segmentation is dilated by 1 pixel. The central mass segmentation is subtracted from the dilated mass segmentation to obtain the central mass border trace.

The spicules feature map is masked by central mass segmentation. The masked result is added to the central mass border to obtain the “hollow” spicules feature map.

Any pixel on the central mass border can be used as a seed, and the flood-fill algorithm is applied over the hollow spicules feature map to obtain the spicules feature segmentation.

The central mass border trace is subtracted from the spicules feature segmentation to obtain the individual spicules segmentation masks.

Those false spicules that do not radiate from the centroid of the central mass segmentation are removed using elongation criteria.

The resulted spicules segmentation mask can be displayed to aid radiologists in cancer diagnosis, or can be used to extract spicules feature for Computer-aided detection server.

FIG. 4 shows an example of the results from the algorithm described in this invention.

An initial mass candidate (410) is selected as region of interest for further processing. The size of the region is defined as 4 times of the estimated mass size obtained from initial candidate calculation. The ROI is then down-sampled to a fixed dimension (such as, 256×256). The pixels outside the breast are filled up by mirroring the pixels inside the breast.

The central mass segmentation (420) is obtained using the algorithm described in FIG. 2.

The central mass border and the spicules feature map (430) are generated using the algorithm described in FIG. 2.

The spicules tracing result (440) is obtained using the algorithm described in FIG. 3.

The tracing result is overlaid on the original image (450) to display on a review workstation to aid radiologists in making diagnosis.

Claims

1. A method to produce the central mass segmentation and the associated spicules trace in digital mammograms, comprising the steps of: selecting mass density candidates to obtain the regions of interest (ROIs);pre-processing each candidate to obtain the central mass segmentation and the corresponding spicules feature map;tracing the central mass border and the associated spicules;displaying the result or using the result in a CAD server.

2. The method of claim 1, wherein said step of pre-processing each candidate to obtain the central mass segmentation and the corresponding spicules feature map comprises the steps of: extracting a region of 4 times of the estimated mass size at each candidate location;re-sampling the region to a fixed dimension, say (256×256);if the region is partially on the breast skin-line or chest-wall edges, the pixels outside the breast are filled up by mirroring the pixels inside the breast;band-pass filtering the region using Gaussian blur difference filter with kernel size 96 and 8;adding the band-pass image weighting by 4 and the directional edge image to form the enhanced image;modulating the image with a “hat” shape modular, said Gaussian weights to obtain the enhanced image.using the flood-fill segmentation technique to the enhanced image to search for central mass segmentation in the range between the seed-pixel-value (center of the candidate ROI) and the maximum-pixel-value within the ROI.applying an adaptive threshold algorithm to the sum of the enhanced image and the edge image to obtain the spicules feature map.

3. The method of claim 1, wherein said step of tracing of the central mass border and the associated spicules comprises the steps of: dilating the central mass segmentation by 1 pixel;subtracting the central mass segmentation from the dilated mass segmentation to obtain the central mass border trace;masking the spicules feature map by central mass segmentation;adding the masked result to the central mass border to obtain the “hollow” spicules feature map;selecting a pixel on the central mass border as a seed;applying the flood-fill algorithm over the hollow spicules feature map to obtain the spicules feature segmentation;subtracting the central mass border trace from the spicules feature segmentation to obtain the individual spicules segmentation masks.removing those false spicules that do not radiate from the centroid of the central mass segmentation using elongation criteria.

4. The method of claim 1, wherein said step of displaying the result includes: displaying the central mass segmentation and the corresponding spicules overlaid on the original image;the overlays are in color rather than gray scale shades on a color display, or adaptive bright (white) color on dark background in original image and dark (black) color on bright background in original image;user interface is provided so the user can toggle on or off the spicules tracing overlays.

5. The method of claim 1, wherein said step of using the result in a CAD server includes: calculating the BIRADS spiculated margin features using the spicules tracing result;the spiculated margin features are (1) a Boolean value indicating if the spicules present; (2) the ratio of the diameter of the central mass to the median length of the spicules; (3) the average angle of the spicules from the center of the central mass.