Systems and methods for classifying an anomaly medical image using variational autoencoder

1. A computer-implemented method for classifying, a medical image, the method comprising:

training a recognition model and a generative model by at least:

inputting a training image into the recognition model, the training image being a normal image;

generating, using the recognition model, one or more training attribute distributions corresponding to the training image;

determining a first loss corresponding to the likelihood of a training sample image deviating from the training image, the training sample image generated by sampling, by the generative model, the one or more training attribute distributions;

determining one or more training distribution divergences based at least in part on comparing the one or more training attribute distributions to a reference distribution;

determining a second loss corresponding to the one or more training distribution divergences; and

changing one or more parameters of the recognition model and one or more parameters of the generative model simultaneously to reduce the sum of the first loss and the second loss;

inputting a medical image into the recognition model, the recognition model configured to:

generate one or more attribute distributions that are substantially Gaussian when inputted with a normal image; and

generate one or more attribute distributions that are substantially non-Gaussian when inputted with an abnormal image;

generating, by the recognition model, one or more attribute distributions corresponding to the medical image;

generating a marginal likelihood corresponding to a likelihood of a sample image substantially matching the medical image, the sample image generated by sampling, by the generative model, the one or more attribute distributions;

comparing the marginal likelihood to a predetermined likelihood threshold; and

generating a classification by at least:

if the marginal likelihood is greater than or equal to the predetermined likelihood threshold, determining the classification of the medical image to be normal; and

if the marginal likelihood is less than the predetermined likelihood threshold, determining the classification of the medical image to be abnormal.

2. The computer-implemented method of claim 1 , wherein the training the recognition model and the generative model is repeated by inputting one or more additional training images.

3. The computer-implemented method of claim 1 , wherein the training image and the medical image are patient images including one or more shared patient features.

4. The computer-implemented method of claim 1 , wherein the one or more training distribution divergences is one or more Kullback-Leibler divergences.

5. The computer-implemented method of claim 1 , wherein the reference distribution is a normal distribution having a mean of zero and a variance of unity.

6. The computer-implemented method of claim 1 , further comprising slicing a three-dimensional training image into a plurality of two-dimensional training images.

7. The computer-implemented method of claim 1 , further comprising compressing the training image into a compressed training image.

8. The computer-implemented method of claim 7 , wherein the compressing the training image includes extracting one or more pertinent features using an autoencoder.

9. The computer-implemented method of claim 7 , wherein the compressing the training image includes reducing the dimensionality of the training image.

10. The computer-implemented method of claim 1 , wherein the recognition model is an encoder and the generative model is a decoder, the encoder and the decoder being part of a variational autoencoder.

11. The computer-implemented method of claim 1 , wherein the generative model is a multi variate Gaussian model with a diagonal covariance matrix.

12. The computer-implemented method of claim 11 , wherein the covariance matrix is an identity matrix.

13. The computer-implemented method of claim 1 , wherein the generating a marginal likelihood includes using the Monte-Carlo sampling method.

14. The computer-implemented method of claim 13 , wherein the Monte-Carlo sampling method utilizes importance sampling.

15. The computer-implemented method of claim 13 , wherein the Monte-Carlo sampling method utilizes Markov chain Monte Carlo.

16. The computer-implemented method of claim 1 , further comprising presenting at least one of the classification and the marginal likelihood to a user.

17. A computer-implemented method for classifying a medical image, the method comprising:

inputting a medical image into the recognition model the recognition model configured to:

generate one or more attribute distributions that are substantially Gaussian when inputted with a normal image; and

generate one or more attribute distributions that are substantially non-Gaussian when inputted with an abnormal image;

generating, by the recognition model, one or more attribute distributions corresponding to the medical image;

generating a marginal likelihood corresponding to a likelihood of a sample image substantially matching the medical image, the sample image generated by sampling, by a generative model, the one or more attribute distributions;

comparing the marginal likelihood to a predetermined likelihood threshold; and

generating a classification by at least:

if the marginal likelihood is greater than or equal to the predetermined likelihood threshold, determining the classification of the medical image to be normal;

if the marginal likelihood is less than the predetermined likelihood threshold, determining the classification of the medical image to be abnormal;

determining one or more distribution divergences based at least in part on comparing the one or more attribute distributions to a reference distribution; and

generating a match likelihood based at least in part on the one or more distribution divergences, the match likelihood corresponding to the likelihood of the sample image substantially matching the medical image;

wherein the higher the match likelihood the more likely the medical image is a normal image rather than an abnormal image;

wherein the lower the match likelihood the more likely the medical image is an abnormal image rather than a normal image.

18. A non-transitory computer-readable medium with instructions stored thereon, that when executed by a processor, causes the processor to perform the processes including:

training a recognition model and a generative model by at least:

inputting a training image into the recognition model, the training image being a normal image;

generating, using the recognition model, one or more training attribute distributions corresponding to the training image;

determining a first loss corresponding to the likelihood of a training sample image deviating from the training image, the training sample image generated by sampling, by the generative model, the one or more training attribute distributions;

determining one or more training distribution divergences based at least in part on comparing the one or more training attribute distributions to a reference distribution;

determining a second loss corresponding to the one or more training distribution divergences; and

changing one or more parameters of the recognition model and one or more parameters of the generative model simultaneously to reduce the sum of the first loss and the second loss;

inputting a medical image into the recognition model, the recognition model configured to:

generate attribute distributions that are substantially Gaussian when inputted with a normal image; and

generate attribute distributions that are substantially non-Gaussian when inputted with an abnormal image;

generating, by the recognition model, one or more attribute distributions corresponding to the medical image;

generating a marginal likelihood corresponding to a likelihood of a sample image substantially matching the medical image, the sample image generated by sampling, by the generative model, the one or more attribute distributions;

comparing the marginal likelihood to a predetermined likelihood threshold; and

generating a classification by at least:

if the marginal likelihood is greater than or equal to the predetermined likelihood threshold, determining the classification of the medical image to be normal; and

if the marginal likelihood is less than the predetermined likelihood threshold, determining the classification of the medical image to be abnormal.