Anomaly detection in medical imaging data

Patent metadata

Field	Value
Publication number	US-12283365-B2
Application number	US-202217695786-A
Country	US
Kind code	B2
Filing date	Mar 15, 2022
Priority date	Mar 16, 2021
Publication date	Apr 22, 2025
Grant date	Apr 22, 2025

Abstract

Particular embodiments provide a method for anomaly detection in images of tissue. An image processing system may receive an image of a tissue sample. A set of tiles may be generated from the image of the tissue sample. The set of tiles may be input into an anomaly detection model comprising a generator model comprising functional skip-connections and a Markovian discriminator model. The anomaly detection model may be trained to isolate a feature space of normal tissue samples. Anomaly scores may be computed for the set of tiles, and an assessment may be generated for the image of the tissue sample based on the anomaly scores for the set of tiles. The assessment may include a reconstructed heatmap of the image of the tissue sample, wherein colors of the heatmap are selected based on the anomaly scores.

First claim

The invention claimed is:

1. A method for anomaly detection, comprising:
receiving, by an image processing system, an image of a tissue sample;
generating, by the image processing system, a set of tiles of the image of the tissue sample;
inputting, by the image processing system, the set of tiles into an anomaly detection model, the anomaly detection model comprising:
a generator model comprising functional skip-connections; and
a Markovian discriminator model;
wherein the anomaly detection model was trained to isolate a feature space of normal tissue samples;

computing, by the image processing system, anomaly scores for the set of tiles; and
generating, by the image processing system and based on the anomaly scores for the set of tiles, an assessment for the image of the tissue sample, wherein the assessment identifies the image of the tissue sample as including abnormal tissue.

2. The method of claim 1 , wherein generating the set of tiles further comprises:
extracting tiles of interest, wherein the set of tiles comprises only tiles of interest.

3. The method of claim 1 , wherein generating the set of tiles comprises:
segmenting the image of the tissue sample into tiles at a plurality of scales of the image, each of the scales representing multiple channels of input.

4. The method of claim 3 , wherein generating the assessment for the image of the tissue sample comprises:
concatenating the plurality of scales of the image to generate a hyper-dimensional image containing multiple levels of information.

5. The method of claim 4 , wherein the concatenated plurality of scales of the image are symmetric about a central pixel along a channel axis.

6. The method of claim 5 , further comprising de-blurring the heatmap image.

7. The method of claim 1 , wherein generating the assessment for the image of the tissue sample comprises:
mapping each unique score of the anomaly scores to a distinct color; and
generating, based on the mapping and the anomaly scores for tiles in the set of tiles, a heatmap image of the image of the tissue sample, wherein each of the tiles in the heatmap image is depicted in the distinct color mapped to the anomaly score for the tile.

8. The method of claim 1 , wherein the computing an anomaly score for the tiles comprises computing an anomaly score for each of a plurality of overlapping patches of the tiles.

9. The method of claim 1 , further comprising normalizing the anomaly scores to fall within a range from 0 to 1.

10. The method of claim 1 , wherein the anomaly score for each of the tiles is computed as a function of a loss of the generator model and a loss of the Markovian discriminator model, wherein the loss of the generator model is determined as between the tile and a corresponding reconstruction of the tile, and wherein the loss of the Markovian discriminator model is determined as between an output of the Markovian discriminator model for the tile and an output of the Markovian discriminator model for the corresponding reconstruction of the tile.

11. The method of claim 10 , wherein training the anomaly detection model is based on a combined loss comprising an adversarial loss, a reconstruction loss, and a latent loss.

12. The method of claim 10 , wherein training the anomaly detection model is based on a reconstruction loss comprising a perceptual loss based on a structural similarity metric.

13. The method of claim 1 , further comprising:
training the anomaly detection model to isolate a feature space of normal tissue samples by using a one-class data set comprising images of normal tissue.

14. The method of claim 1 , wherein the functional skip-connections are regularized by randomly dropping at least one node from the generator network for at least one of the training tiles.

15. One or more computer-readable non-transitory storage media embodying software comprising instructions operable when executed to:
receive, by an image processing system, an image of a tissue sample;
generate, by the image processing system, a set of tiles of the image of the tissue sample;
input, by the image processing system, the set of tiles into an anomaly detection model, the anomaly detection model comprising:
a generator model comprising functional skip-connections; and
a Markovian discriminator model;
wherein the anomaly detection model was trained to isolate a feature space of normal tissue samples;

compute, by the image processing system, anomaly scores for the set of tiles; and
generate, by the image processing system and based on the anomaly scores for the set of tiles, an assessment for the image of the tissue sample, wherein the assessment identifies the image of the tissue sample as including abnormal tissue.

16. The computer-readable non-transitory storage media of claim 15 , wherein instructions operable when executed to generate the assessment for the image of the tissue sample comprises instructions operable when executed to:
map each unique score of the anomaly scores to a distinct color; and
generate, based on the mapping and the anomaly scores for tiles in the set of tiles, a heatmap image of the image of the tissue sample, wherein each of the tiles in the heatmap image is depicted in the distinct color mapped to the anomaly score for the tile.

17. The computer-readable non-transitory storage media of claim 15 , the software further comprising instructions operable when executed to:
train the anomaly detection model to isolate a feature space of normal tissue samples by using a one-class data set comprising images of normal tissue.

18. A image processing system comprising one or more processors and a memory coupled to the processors comprising instructions executable by the processors, the processors being operable when executing the instructions to:
receive an image of a tissue sample;
generate a set of tiles of the image of the tissue sample;
input the set of tiles into an anomaly detection model, the anomaly detection model comprising:
a generator model comprising functional skip-connections; and
a Markovian discriminator model;
wherein the anomaly detection model was trained to isolate a feature space of normal tissue samples;

compute anomaly scores for the set of tiles; and
generate, based on the anomaly scores for the set of tiles, an assessment for the image of the tissue sample, wherein the assessment identifies the image of the tissue sample as including abnormal tissue.

19. The image processing system of claim 18 , wherein the processors being operable when executing the instructions to generate the assessment for the image of the tissue sample comprises the processors being operable when executing the instructions to:
map each unique score of the anomaly scores to a distinct color; and
generate, based on the mapping and the anomaly scores for tiles in the set of tiles, a heatmap image of the image of the tissue sample, wherein each of the tiles in the heatmap image is depicted in the distinct color mapped to the anomaly score for the tile.

20. The image processing system of claim 18 , the processors being further operable when executing the instructions to:
train the anomaly detection model to isolate a feature space of normal tissue samples by using a one-class data set comprising images of normal tissue.

Assignees

• Genentech Inc

Inventors

• Hu fang-yao

Classifications

• G06N20/00

  Machine learning · CPC title

• G06N3/084

  Backpropagation, e.g. using gradient descent · CPC title

• G06N3/082

  modifying the architecture, e.g. adding, deleting or silencing nodes or connections · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/0455

  Auto-encoder networks; Encoder-decoder networks · CPC title