Auto-normalization for machine learning
Systems and methods for using a prediction model jointly with a normalization model to provide prediction results are provided. One example method includes receiving an input image of a tissue sample of a patient and generating a normalized image by applying a normalization model on the input image. The normalization model is configured to generate normalized data using input data for a prediction model, and the prediction model is configured to generate prediction results using normalized data generated by the normalization model. The normalization model and the prediction model are jointly trained. The method further includes generating a prediction of disease severity for the patient by applying the prediction model on the normalized image.
1. A computer-implemented method, comprising:
receiving an input image of a tissue sample of a patient;
generating a normalized image by applying a normalization model on the input image, wherein the normalization model is configured to generate normalized data using input data for a prediction model, the prediction model configured to generate prediction results using normalized data generated by the normalization model, and wherein the normalization model and the prediction model are jointly trained by comparing a first set of prediction results generated by the prediction model using a first set of normalized training inputs generated by applying the normalization model to training inputs in a set of training samples for a first number of times and a second set of prediction results generated by the prediction model using a second set of normalized training inputs generated by applying the normalization model to the training inputs for a second number of times, the second number of times being different from the first number of times; and
generating a prediction for the patient by applying the prediction model on the normalized image.
2. The method of claim 1 , wherein the normalization model and the prediction model are jointly trained based on a loss function comprising an idempotence loss term measuring a difference between the first set of prediction results and the second set of prediction results.
3. The method of claim 2 , wherein the set of training samples comprise unlabeled samples that comprise the training inputs without corresponding training outputs, and wherein the idempotence loss term is calculated based on the training inputs in the unlabeled samples.
4. The method of claim 3 , wherein the set of training samples further comprises labeled samples that comprise training inputs and corresponding training outputs, and wherein the loss function further comprises a loss term for the labeled samples measuring a difference between prediction results generated by the prediction model and the normalization model based on the training inputs and the corresponding training outputs.
5. The method of claim 2 , wherein the idempotence loss term comprises a Kullback-Leibler divergence between the first set of prediction results and the second set of prediction results, a mean absolute error between the first set of prediction results and the second set of prediction results, or a mean square error between the first set of prediction results and the second set of prediction results.
6. The method of claim 2 , wherein the loss function further comprises an entropy loss term representing an entropy of the first set of prediction results or the second set of prediction results.
7. The method of claim 6 , wherein the input image is a pathology image and generating the prediction for the patient comprises detecting presence or absence of metastatic lesions in the pathology image using the prediction model.
8. A system comprising:
at least one processor; and
at least one non-transitory computer-readable medium comprising processor-executable instructions stored thereupon, which, when executed by the at least one processor, cause the processor to:
receive an input image of a tissue sample of a patient;
generate a normalized image by applying a normalization model on the input image, wherein the normalization model is configured to generate normalized data using input data for a prediction model, the prediction model configured to generate prediction results using normalized data generated by the normalization model, and wherein the normalization model and the prediction model are jointly trained by comparing a first set of prediction results generated by the prediction model using a first set of normalized training inputs generated by applying the normalization model to training inputs in a set of training samples for a first number of times and a second set of prediction results generated by the prediction model using a second set of normalized training inputs generated by applying the normalization model to the training inputs for a second number of times, the second number of times being different from the first number of times; and
generate a prediction for the patient by applying the prediction model on the normalized image.
9. The system of claim 8 , wherein the normalization model and the prediction model are jointly trained based on a loss function comprising an idempotence loss term measuring a difference between the first set of prediction results and the second set of prediction results.
10. The system of claim 9 , wherein the set of training samples comprise unlabeled samples that comprise the training inputs without corresponding training outputs, and wherein the idempotence loss term is calculated based on the training inputs in the unlabeled samples.
11. The system of claim 10 , wherein the set of training samples further comprises labeled samples that comprise training inputs and corresponding training outputs, and wherein the loss function further comprises a loss term for the labeled samples measuring a difference between prediction results generated by the prediction model and the normalization model based on the training inputs and the corresponding training outputs.
12. The system of claim 9 , wherein the idempotence loss term comprises a Kullback-Leibler divergence between the first set of prediction results and the second set of prediction results, a mean absolute error between the first set of prediction results and the second set of prediction results, or a mean square error between the first set of prediction results and the second set of prediction results.
13. The system of claim 9 , wherein the loss function further comprises an entropy loss term representing an entropy of the first set of prediction results or the second set of prediction results.
14. The system of claim 13 , wherein the input image is a pathology image and generating the prediction for the patient comprises detecting presence or absence of metastatic lesions in the pathology image using the prediction model.
15. A non-transitory computer-readable medium comprising processor-executable instructions to cause a processor to:
receive an input image of a tissue sample of a patient;
generate a normalized image by applying a normalization model on the input image, wherein the normalization model is configured to generate normalized data using input data for a prediction model, the prediction model configured to generate prediction results using normalized data generated by the normalization model, and wherein the normalization model and the prediction model are jointly trained by comparing a first set of prediction results generated by the prediction model using a first set of normalized training inputs generated by applying the normalization model to training inputs in a set of training samples for a first number of times and a second set of prediction results generated by the prediction model using a second set of normalized training inputs generated by applying the normalization model to the training inputs for a second number of times, the second number of times being different from the first number of times; and
generate a prediction for the patient by applying the prediction model on the normalized image.
16. The non-transitory computer-readable medium of claim 15 , wherein the normalization model and the prediction model are jointly trained based on a loss function comprising an idempotence loss term measuring a difference between the first set of prediction results and the second set of prediction results.
17. The non-transitory computer-readable medium of claim 16 , wherein the set of training samples comprise unlabeled samples that comprise the training inputs without corresponding training outputs, and wherein the idempotence loss term is calculated based on the training inputs in the unlabeled samples.
18. The non-transitory computer-readable medium of claim 17 , wherein the set of training samples further comprises labeled samples that comprise training inputs and corresponding training outputs, and wherein the loss function further comprises a loss term for the labeled samples measuring a difference between prediction results generated by the prediction model and the normalization model based on the training inputs and the corresponding training outputs.
19. The non-transitory computer-readable medium of claim 16 , wherein the idempotence loss term comprises a Kullback-Leibler divergence between the first set of prediction results and the second set of prediction results, a mean absolute error between the first set of prediction results and the second set of prediction results, or a mean square error between the first set of prediction results and the second set of prediction results.
20. The non-transitory computer-readable medium of claim 16 , wherein the loss function further comprises an entropy loss term representing an entropy of the first set of prediction results or the second set of prediction results.