Malignant mass detection and classification in radiographic images

• US 9,262,822 B2
• Filed: 04/29/2011
• Issued: 02/16/2016
• Est. Priority Date: 04/30/2010
• Status: Active Grant

• Current Assignee: iCAD Incorporated
• Original Assignee: VuCOMP, Inc.
• Inventors: Wehnes, Jeffrey C., Pike, James H.
• Primary Examiner(s): FUJITA, KATRINA R
• Application Number:
  US13/695,357
• Publication Number:
  US 20130202165A1
• Time in Patent Office: 1,754 Days
• US Class Current: 1/1
• CPC Class Codes: G06F 18/24 Classification techniques
• CPC Class Codes: G06T 2207/10116 X-ray image
• CPC Class Codes: G06T 2207/20016 Hierarchical, coarse-to-fine, multiscale or multiresolution image processing; Pyramid transform
• CPC Class Codes: G06T 2207/30068 Mammography; Breast
• CPC Class Codes: G06T 2207/30096 Tumor; Lesion
• CPC Class Codes: G06T 5/10 using non-spatial domain filtering
• CPC Class Codes: G06T 5/70 Denoising; Smoothing
• CPC Class Codes: G06T 7/0012 Biomedical image inspection
• CPC Class Codes: G06T 7/11 Region-based segmentation
• CPC Class Codes: G06T 7/143 involving probabilistic approaches, e.g. Markov random field [MRF] modelling
• CPC Class Codes: G06V 10/454 Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]

Abstract

An image analysis embodiment comprises subsampling a digital image by a subsample factor related to a first anomaly size scale, thereby generating a subsampled image, smoothing the subsampled image to generate a smoothed image, determining a minimum negative second derivative for each pixel in the smoothed image, determining each pixel having a convex down curvature based on a negative minimum negative second derivative value for the respective pixel, joining each eight-neighbor connected pixels having convex down curvature to identify each initial anomaly area, selecting the initial anomaly areas having strongest convex down curvatures based on a respective maximum negative second derivative for each of the initial anomaly areas, extracting one or more classification features for each selected anomaly area, and classifying the selected anomaly areas based on the extracted one or more classification features.

Claims

1. A method for processing an image comprising pixels, the method comprising:

identifying anomalies in the image at a first anomaly size scale, comprising;

subsampling a digital image by a subsample factor related to the first anomaly size scale, thereby generating a subsampled image having a lower resolution than the digital image in accordance with the subsample factor;

smoothing the subsampled image to generate a smoothed image;

determining a least-positive negated-second derivative for each pixel in the smoothed image;

determining each pixel having a convex down curvature based on a least-positive negated-second derivative value for the respective pixel;

joining each eight-neighbor connected pixels having convex down curvature to identify each initial anomaly area having only the first anomaly size scale, wherein, for each initial anomaly area, an internal pixel of the respective initial anomaly area represents a point of least curvature for the respective initial anomaly area;

selecting one or more of the initial anomaly areas having only the first anomaly size scale and having strongest convex down curvatures based on a respective maximum least-positive negated-second derivative value at the respective internal pixel for each of the initial anomaly areas;

extracting one or more classification features for each selected anomaly area having only the first anomaly size scale; and

classifying the selected anomaly areas having only the first anomaly size scale based on the extracted one or more classification features; and

performing the identifying the anomalies in the image for a plurality of different anomaly size scales.

2. The method of claim 1, wherein the image is a mammogram, the anomalies are potentially-malignant masses, and the one or more classification features are selected from the group consisting of:

search width index, nipple distance and y position, signal-to-noise ratio (SNR), object rank, relative arc length, dip SNR, global SNR, other side SNR, and combinations thereof.

3. The method of claim 1, further comprising, before subsampling, removing bright areas from the digital image.

4. The method of claim 3, further comprising, after removing the bright areas and before subsampling, flattening an intensity of the digital image.

5. The method of claim 1, wherein determining the least-positive negated-second derivative for each pixel comprises:

calculating second derivatives at three image points centered at each pixel and linearly spaced at a large scale; and

repeating calculating second derivatives for a plurality of orientations around the each pixel.

6. The method of claim 1, further comprising building a noise map of the smoothed image by:

calculating second derivatives at three image points centered at each pixel and linearly spaced at a fine scale;

repeating calculating second derivatives for a plurality of orientations around the each pixel; and

using a minimum absolute second derivative from the plurality of orientations as a noise map value for the each pixel.

7. The method of claim 1, further comprising, after joining and before selecting, eroding and dilating each initial anomaly area.

8. The method of claim 1, further comprising, before extracting, refining a boundary of each selected initial anomaly area.

9. A system for identifying anomalies in an image comprising pixels, the system comprising:

an image subsampler subsampling a digital image using a subsample factor related to a first anomaly size scale, generating a subsampled image having a lower resolution than the digital image in accordance with the subsample factor;

an image smoother smoothing the subsampled image, generating a smoothed image;

a curvature signature detector;

determining a least-positive negated-second derivative for each pixel in the smoothed image,

detecting convex down curvature based on a least-positive negated-second derivative value for each pixel in the smoothed image, and

detecting and joining neighboring convex down curvatures in the smoothed image to generate anomaly areas having only the first anomaly size scale, wherein, for each anomaly area, an internal pixel of the respective anomaly area represents a point of least curvature for the respective anomaly area;

an anomaly selector selecting one or more of the anomaly areas having only the first anomaly size scale and having strongest convex down curvatures based on a respective maximum least-positive negated-second derivative value at the respective internal pixel for each of the anomaly areas;

a feature extractor extracting one or more classification features for each of the selected anomaly areas having only the first anomaly size scale; and

a classifier classifying the selected anomaly areas having only the first anomaly size scale based on one or more thresholds for the extracted one or more classification features,

wherein the image subsampler, the image smoother, the curvature signature detector, the anomaly selector, the feature extractor and the classifier perform their respective functions for the image for a plurality of different anomaly size scales.

10. The system of claim 9, wherein the image is a mammogram, the anomalies are potentially-malignant masses, and the one or more classification features are selected from the group consisting of:

search width index, nipple distance and y position, signal-to-noise ratio (SNR), object rank, relative arc length, dip SNR, global SNR, other side SNR, and combinations thereof.

11. The system of claim 9, further comprising the system removing bright areas from the digital image, before the image subsampler subsampling the digital image.

12. The system of claim 11, further comprising the system flattening an intensity of the digital image, after the bright areas are removed and before the image subsampler subsampling the digital image.

13. The system of claim 9, wherein the curvature detector determining the least-positive negated-second derivative for each pixel comprises the curvature detector:

calculating second derivatives at three image points centered at each pixel and linearly spaced at a large scale; and

repeating calculating second derivatives for a plurality of orientations around the each pixel.

14. The system of claim 9, further comprising the system building a noise map of the smoothed image by:

calculating second derivatives at three image points centered at each pixel and linearly spaced at a fine scale;

repeating calculating second derivatives for a plurality of orientations around the each pixel; and

using a minimum absolute second derivative from the plurality of orientations as a noise map value for the each pixel.

15. The system of claim 9, further comprising the system eroding and dilating each initial anomaly area, after the curvature signature detector joining and before the anomaly selector selecting.

16. The system of claim 9, further comprising the system refining a boundary of each selected initial anomaly area, before the feature extractor extracting.

17. A computer program product for processing an image, the computer program product having a non-transitory computer-readable medium with a computer program embodied thereon, the computer program comprising:

computer program code for identifying anomalies in the image at a first anomaly size scale, comprising;

computer program code for subsampling a digital image by a subsample factor related to a first anomaly size scale, thereby generating a subsampled image having a lower resolution than the digital image in accordance with the subsample factor;

computer program code for smoothing the subsampled image to generate a smoothed image;

computer program code for determining a least-positive negated-second derivative for each pixel in the smoothed image;

computer program code for determining each pixel having a convex down curvature based on a least-positive second derivative value for the respective pixel;

computer program code for joining each eight-neighbor connected pixels having convex down curvature to identify each initial anomaly area having only the first anomaly size scale, wherein, for each initial anomaly area, an internal pixel of the respective initial anomaly area represents a point of least curvature for the respective initial anomaly area;

computer program code for selecting one or more of the initial anomaly areas having only the first anomaly size scale and having strongest convex down curvatures based on a respective maximum least-positive negated-second derivative value for each of the initial anomaly areas;

computer program code for extracting one or more classification features for each selected anomaly area having only the first anomaly size scale; and

computer program code for classifying the selected anomaly areas having only the first anomaly size scale based on the extracted one or more classification features; and

computer program code for performing the identifying the anomalies in the image for a plurality of different anomaly size scales.

18. The computer program product of claim 17, wherein the image is a mammogram, the anomalies are potentially-malignant masses, and the one or more classification features are selected from the group consisting of:

search width index, nipple distance and y position, signal-to-noise ratio (SNR), object rank, relative arc length, dip SNR, global SNR, other side SNR, and combinations thereof.

19. The computer program product of claim 17, further comprising computer program code for removing bright areas from the digital image prior to subsampling.

20. The computer program product of claim 17, wherein the computer program code for determining the least positive negated-second derivative for each pixel comprises:

computer program code for calculating second derivatives at three image points centered at each pixel and linearly spaced at a large scale; and

computer program code for repeating calculating second derivatives for a plurality of orientations around the each pixel.

21. The computer program product of claim 19, further comprising computer program code for flattening an intensity of the digital image after removing the bright areas and before subsampling.

22. The computer program product of claim 17, further comprising computer program code for building a noise map of the smoothed image, comprising:

computer program code for calculating second derivatives at three image points centered at each pixel and linearly spaced at a fine scale;

computer program code for repeating calculating second derivatives for a plurality of orientations around the each pixel; and

computer program code for using a minimum absolute second derivative from the plurality of orientations as a noise map value for the each pixel.

23. The computer program product of claim 17, further comprising computer program code for eroding and dilating each initial anomaly area after joining and before selecting.

24. The computer program product of claim 17, further comprising computer program code for refining a boundary of each selected initial anomaly area before extracting.