Anomaly detection in medical imagery

Patent metadata

Field	Value
Publication number	US-2017148166-A1
Application number	US-201715427073-A
Country	US
Kind code	A1
Filing date	Feb 8, 2017
Priority date	Feb 12, 2014
Publication date	May 25, 2017
Grant date	—

Abstract

A method comprising using at least one hardware processor for computing a patch distinctiveness score for each of multiple patches of a medical image, computing a shape distinctiveness score for each of multiple regions of the medical image, and computing a saliency map of the medical image, by combining the patch distinctiveness score and the shape distinctiveness score.

First claim

What is claimed is:

1 . A method comprising using at least one hardware processor for:
computing a patch distinctiveness score for each of multiple patches of a medical image; computing a shape distinctiveness score for each of multiple regions of the medical image, wherein said computing of the shape distinctiveness score comprises:
applying an edge detection algorithm to each of the multiple regions, to detect at least one pair of boundary edges in each of at least some of the multiple regions,
for each pair of boundary edges (p, q):
(a) computing a length (l pq ) of a vector ({right arrow over (pq)}) and an orientation (ψ pq ) of the vector ({right arrow over (pq)}),
(b) computing a normal (θ p ) to the boundary edge (p) and a normal (θ q ) to the boundary edge (q), and
(c) computing histograms for l pq , ψ pq , θ p and θ q , and

computing the shape distinctiveness score for each of the at least some of the multiple regions, based on an entropy computation of the histograms; and
computing a saliency map of the medical image, by combining the patch distinctiveness score and the shape distinctiveness score.

2 . The method according to claim 1 , wherein said computing of the patch distinctiveness score comprises:
applying principal component analysis (PCA) to the multiple patches, to represent each of the multiple patches by a set of expansion coefficients; and summing the set of expansion coefficients, to produce the patch distinctiveness score for each of the multiple patches.

3 . The method according to claim 1 , wherein said computing of the shape distinctiveness score further comprises, following step (c), normalizing the histograms.

4 . The method according to claim 1 , wherein each of the histograms comprises k bins, and wherein the entropy is computed by:
H ( h f )=−Σ k h f ( k )log( h f ( k )).

5 . The method according to claim 4 , wherein the shape distinctiveness score for each of the at least some of the multiple regions (R) is defined by:
SD ( R )=⅓( H ( h ψ )+ H ( h θ p )+ H ( h θ q ))− H ( h l ).

6 . The method according to claim 1 , wherein said combining of the patch distinctiveness score and the shape distinctiveness score comprises:
decomposing the medical image to connected components based on the multiple regions shape distinctiveness score; and in each connected component, multiplying the patch distinctiveness score at the location of the maximal shape distinctiveness score by a Gaussian.

7 . The method according to claim 6 , further comprising using said at least one hardware processor for:
applying a mode seeking algorithm to the saliency map, to detect one or more salient regions; and applying a medical condition classifier to the salient regions, to classify the one or more salient regions.

8 . A method comprising:
repeating the method of claim 1 for a plurality of medical images, to produce a plurality of saliency maps, wherein the plurality of medical images comprises images of normal and abnormal medical conditions; detecting salient regions in the plurality of saliency maps; classifying the salient regions by using a classifier to produce a medical condition classifier; conducting the method of claim 1 for an additional medical image, to receive a saliency map of the additional medical image; detecting one or more salient regions in the saliency map; and employing the medical condition classifier on the one or more salient regions of the additional medical image, to determine if the medical image is of a normal or an abnormal medical condition.

9 . The method according to claim 8 , wherein the plurality of medical images and the additional medical image are mammograms, and the abnormal medical condition is selected from the group consisting of: a microcalcifications and a tumor.

10 . The method according to claim 8 , wherein the plurality of medical images and the additional medical image are angiograms, and the abnormal medical condition is blood vessel stenosis.

11 . The method according to claim 8 , wherein the plurality of medical images and the additional medical image are MRI (magnetic resonance imaging) images, and the abnormal medical condition is a lesion.

12 . The method according to claim 8 , wherein the plurality of medical images and the additional medical image are CT (computed tomography) images, and the abnormal medical condition is a lesion.

13 . The method according to claim 8 , wherein the plurality of medical images and the additional medical image are ultrasound images, and the abnormal medical condition is a lesion.

14 . The method according to claim 1 , wherein a division of the medical image into the multiple patches is different than a division of the medical image into the multiple regions.

15 . A computer program product comprising a non-transitory computer-readable storage medium having program code embodied therewith, the program code executable by at least one hardware processor to:
compute a patch distinctiveness score for each of multiple patches of a medical image; compute a shape distinctiveness score for each of multiple regions of the medical image, wherein said computing of the shape distinctiveness score comprises:
applying an edge detection algorithm to each of the multiple regions, to detect at least one pair of boundary edges in each of at least some of the multiple regions,
for each pair of boundary edges (p, q):
(a) computing a length (l pq ) of a vector ({right arrow over (pq)}) and an orientation (ψ pq ) of the vector ({right arrow over (pq)}),
(b) computing a normal (θ p ) to the boundary edge (p) and a normal (θ q ) to the boundary edge (q), and
(c) computing histograms for l pq , ψ pq , θ p and θ q , and

computing the shape distinctiveness score for each of the at least some of the multiple regions, based on an entropy computation of the histograms; and
compute a saliency map of the medical image, by combining the patch distinctiveness score and the shape distinctiveness score.

16 . The computer program product according to claim 15 , wherein the computation of the patch distinctiveness score comprises:
applying principal component analysis (PCA) to the multiple patches, to represent each of the multiple patches by a set of expansion coefficients; and summing the set of expansion coefficients, to produce the patch distinctiveness score for each of the multiple patches.

17 . The computer program product according to claim 15 , wherein the program code is further executable by said at least one hardware processor to:
apply a mode seeking algorithm to the saliency map, to detect one or more salient regions; and apply a medical condition classifier to the one or more salient regions, to classify the one or more salient regions.

18 . The computer program product of claim 15 , wherein the computation of the patch distinctiveness score, the shape distinctiveness score and the saliency map is performed for a plurality of medical images, to produce a plurality of saliency maps, wherein the plurality of medical images comprise images of normal and abnormal medical conditions, and wherein the program code is further executable by said at least one hardware processor to:
detect salient regions in the plurality of saliency maps; classify the salient regions by using a classifier to produce a medical condition classifier; conduct the computation of the patch distinctiveness score, the shape distinctiveness score and the saliency map for an additional medic

Assignees

• IBM

Inventors

• Alpert Sharon
• Kisilev Pavel

Classifications

• G06V10/462

  Salient features, e.g. scale invariant feature transforms [SIFT] · CPC title

• G06T7/0012 Primary

  Biomedical image inspection · CPC title

• G06F18/24

  Classification techniques · CPC title

• G06V10/469

  Contour-based spatial representations, e.g. vector-coding · CPC title

• G06V10/50

  by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis · CPC title