Systems and methods for automatic detection of an indication of abnormality in an anatomical image

Patent metadata

Field	Value
Publication number	US-10878569-B2
Application number	US-201815937884-A
Country	US
Kind code	B2
Filing date	Mar 28, 2018
Priority date	Mar 28, 2018
Publication date	Dec 29, 2020
Grant date	Dec 29, 2020

Abstract

There is provided a method for training a deep convolutional neural network (CNN) for detecting an indication of likelihood of abnormality, comprising: receiving anatomical training images, each including an associated annotation indicative of abnormality for the whole image without an indication of location of the abnormality, executing, for each anatomical training image: decomposing the anatomical training image into patches, computing a feature representation of each patch, computing for each patch, according to the feature representation of the patch, a probability that the patch includes an indication of abnormality, setting a probability indicative of likelihood of abnormality in the anatomical image according to the maximal probability value computed for one patch, and training a deep CNN for detecting an indication of likelihood of abnormality in a target anatomical image according to the patches of the anatomical training images, the one patch, and the probability set for each respective anatomical training image.

First claim

What is claimed is:

1. A computer implemented method for training a deep convolutional neural network (CNN) for detecting an indication of likelihood of abnormality in a target anatomical image based on a plurality of anatomical training images each associated with an annotation for a whole respective training image, comprising:
receiving the plurality of anatomical training images, each including an associated annotation indicative of abnormality for the whole respective anatomical training image without an indication of a location of the abnormality within the respective anatomical image;
executing, for each respective anatomical training image of the plurality of anatomical training images:
decomposing the respective anatomical training image into a plurality of patches;
computing a feature representation of each patch of the plurality of patches;
computing for each respective patch of the plurality of patches, according to the feature representation of the respective patch, a probability that the respective patch includes an indication of abnormality;
setting a probability indicative of likelihood of abnormality in the respective anatomical image according to the maximal probability value computed for one patch of the plurality of patches; and

training the deep convolutional neural network for detecting the indication of likelihood of abnormality in the target anatomical image according to the plurality of patches of the plurality of anatomical training images, the one patch, and the probability set for each respective anatomical training image;
wherein the deep CNN is trained according to a loss function that computes a log likelihood loss according to a probability that a certain patch of the plurality of patches is classified as indicative of abnormality based on a plurality of coefficients of the deep CNN; and
wherein the loss function is mathematically represented as:

⁢

(
θ
)

=

∑

X
i

∈
Λ

⁢

⁢

Y
i

=

y
+

⁢

⁢

log
(

max

x
ij

∈

X
i

⁢

⁢

(

y
+

|

x
ij

,
θ

)

)

+

∑

X
i

∈
Λ

⁢

⁢

Y
i

=

y
-

⁢

log
(

1
-

max

x
ij

∈

X
i

⁢

(

⁢

(

y
+

|

x
ij

,
θ

)

)

wherein:
x ji denotes the respective patch of the respective anatomical image,
θ denotes the coefficients of the deep CNN,

(y+|x ij ,θ) denotes a probability that the respective patch denoted x ji is classified as positive based on the coefficients θ of the deep CNN.

2. The method according to claim 1 , wherein an abnormality appearing in each one of the plurality of anatomical training images is not associated with a manual annotation indicative of location of the abnormality within the respective anatomical training image.

3. The method according to claim 1 , wherein the loss function considers the one patch of the plurality of patches most probably indicative of abnormality and excludes other patches of the plurality of patches with lower probability values than the one patch, wherein the one patch is back propagated through the deep CNN for updating of the coefficients of the deep CNN.

4. The method according to claim 3 , wherein the probability comprises a probabilistic geometric prior value denoting areas on a border of at least one tissue portion based on distance from an edge of the

Assignees

• IBM

Inventors

• Akselrod-Ballin Ayelet
• Bakalo Ran
• Ben-Ari Rami
• Choukroun Yoni
• Kisilev Pavel

Classifications

• G16H30/40 Primary

  for processing medical images, e.g. editing · CPC title

• G06V10/82

  using neural networks · CPC title

• G06V10/454

  Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN] · CPC title

• G06V10/764

  using classification, e.g. of video objects · CPC title

• G06T7/0014 Primary

  using an image reference approach · CPC title