Knockout autoencoder for detecting anomalies in biomedical images

1. A method, in a data processing system comprising a processor and a memory, the memory comprising instructions that are executed by the processor to specifically configure the processor to implement a knockout autoencoder engine for detecting anomalies in biomedical images, the method comprising:

training a neural network to be used as a knockout autoencoder that predicts an original image based on an input image, wherein training the neural network comprises:

for each training image in a set of training images, selecting one or more knockout patches of the training image, filling the one or more knockout patches with noise to form a knockout image, and training the neural network to predict the training image based on expected content in the knockout patches of the knockout image given remaining pixels in the knockout image;

providing, by the knockout autoencoder engine, a biomedical image as the input image to the neural network;

outputting, by the neural network, a probability distribution for each pixel in the biomedical image, wherein each probability distribution represents a predicted probability distribution of expected pixel values for a given pixel in the biomedical image;

determining, by an anomaly detection component executing within the knockout autoencoder engine, a probability that each pixel has an expected value based on the probability distributions to form a plurality of expected pixel probabilities;

detecting, by the anomaly detection component, an anomaly in the biomedical image based on the plurality of expected pixel probabilities;

marking, by an anomaly marking component executing within the knockout autoencoder engine, the detected anomaly in the biomedical image to form a marked biomedical image; and

outputting, by the knockout autoencoder engine, the marked biomedical image.

2. The method of claim 1 , wherein the set of training images comprise a plurality of biomedical images of healthy patients.

3. The method of claim 1 , wherein selecting one or more knockout patches comprises selecting one or more random patches in the training image.

4. The method of claim 1 , wherein the neural network comprises a convolutional neural network.

5. The method of claim 4 , wherein the convolutional neural network comprises a U-net.

6. The method of claim 1 , wherein determining the probability that each pixel has an expected value based on the probability distributions comprises:

for each given pixel in the biomedical image having a given pixel value, identifying a probability value corresponding to the given pixel value in a given probability distribution corresponding to the given pixel.

7. The method of claim 1 , wherein detecting the anomaly in the biomedical image based on the plurality of expected pixel probabilities comprises:

comparing each expected pixel probability in the plurality of expected pixel probabilities to a predetermined threshold; and

responsive to one or more of the plurality of expected pixel probabilities being less than the predetermined threshold, determining that the biomedical image has an anomaly.

8. A computer program product comprising a non-transitory computer readable medium having a computer readable program stored therein, wherein the computer readable program, when executed on a computing device, causes the computing device to implement a knockout autoencoder engine for detecting anomalies in biomedical images, wherein the computer readable program causes the computing device to:

train a neural network to be used as a knockout autoencoder that predicts an original image based on an input image, wherein training the neural network comprises:

for each training image in a set of training images, selecting one or more knockout patches of the training image, filling the one or more knockout patches with noise to form a knockout image, and training the neural network to predict the training image based on expected content in the knockout patches of the knockout image given remaining pixels in the knockout image;

provide, by the knockout autoencoder engine, a biomedical image as the input image to the neural network;

output, by the neural network, a probability distribution for each pixel in the biomedical image, wherein each probability distribution represents a predicted probability distribution of expected pixel values for a given pixel in the biomedical image;

determine, by an anomaly detection component executing within the knockout autoencoder engine, a probability that each pixel has an expected value based on the probability distributions to form a plurality of expected pixel probabilities;

detect, by the anomaly detection component, an anomaly in the biomedical image based on the plurality of expected pixel probabilities;

mark, by an anomaly marking component executing within the knockout autoencoder engine, the detected anomaly in the biomedical image to form a marked biomedical image; and

output, by the knockout autoencoder engine, the marked biomedical image.

9. The computer program product of claim 8 , wherein set of training images comprise a plurality of biomedical images of healthy patients.

10. The computer program product of claim 8 , wherein selecting one or more knockout patches comprises selecting one or more random patches in the training image.

11. The computer program product of claim 8 , wherein the neural network comprises a convolutional neural network.

12. The computer program product of claim 11 , wherein the convolutional neural network comprises a U-net.

13. The computer program product of claim 8 , wherein determining the probability that each pixel has an expected value based on the probability distributions comprises:

for each given pixel in the biomedical image having a given pixel value, identifying a probability value corresponding to the given pixel value in a given probability distribution corresponding to the given pixel.

14. The computer program product of claim 8 , wherein detecting the anomaly in the biomedical image based on the plurality of expected pixel probabilities comprises:

comparing each expected pixel probability in the plurality of expected pixel probabilities to a predetermined threshold; and

responsive to one or more of the plurality of expected pixel probabilities being less than the predetermined threshold, determining that the biomedical image has an anomaly.

15. An apparatus, comprising:

a processor; and

a memory coupled to the processor, wherein the memory comprises instructions which, when executed by the processor, cause the processor to implement a knockout autoencoder engine for detecting anomalies in biomedical images, wherein the instructions cause the processor to:

train a neural network to be used as a knockout autoencoder that predicts an original image based on an input image, wherein training the neural network comprises:

for each training image in a set of training images, selecting one or more knockout patches of the training image, filling the one or more knockout patches with noise to form a knockout image, and training the neural network to predict the training image based on expected content in the knockout patches of the knockout image given remaining pixels in the knockout image;

provide, by the knockout autoencoder engine, a biomedical image as the input image to the neural network;

output, by the neural network, a probability distribution for each pixel in the biomedical image, wherein each probability distribution represents a predicted probability distribution of expected pixel values for a given pixel in the biomedical image;

determine, by an anomaly detection component executing within the knockout autoencoder engine, a probability that each pixel has an expected value based on the probability distributions to form a plurality of expected pixel probabilities;

detect, by the anomaly detection component, an anomaly in the biomedical image based on the plurality of expected pixel probabilities;

mark, by an anomaly marking component executing within the knockout autoencoder engine, the detected anomaly in the biomedical image to form a marked biomedical image; and

output, by the knockout autoencoder engine, the marked biomedical image.

16. The apparatus of claim 15 , wherein determining the probability that each pixel has an expected value based on the probability distributions comprises:

for each given pixel in the biomedical image having a given pixel value, identifying a probability value corresponding to the given pixel value in a given probability distribution corresponding to the given pixel.

17. The apparatus of claim 15 ; wherein detecting the anomaly in the biomedical image based on the plurality of expected pixel probabilities comprises:

comparing each expected pixel probability in the plurality of expected pixel probabilities to a predetermined threshold; and

responsive to one or more of the plurality of expected pixel probabilities being less than the predetermined threshold, determining that the biomedical image has an anomaly.

18. The apparatus of claim 15 , wherein the set of training images comprise a plurality of biomedical images of healthy patients.

19. The apparatus of claim 15 , wherein selecting one or more knockout patches comprises selecting one or more random patches in the training image.

20. The apparatus of claim 15 , wherein the neural network comprises a convolutional neural network.