SELF-SUPERVISED TRAINING AT SCALE WITH WEAKLY-SUPERVISED LATENT SPACE STRUCTURE
Systems and methods for performing a medical analysis task using a trained machine learning based task network are provided. Input medical data is received. A medical analysis task is performed using a trained machine learning based task network based on the input medical data. Results of the medical analysis task are output. The trained machine learning based task network is trained by: receiving unannotated training medical data; generating weakly-supervised labels for the unannotated training medical data using one or more trained machine learning based supervised learning networks; training the machine learning based task network for performing the medical analysis task based on 1) the unannotated training medical data, 2) self-supervised labels for the unannotated training medical data learned via self-supervised learning, and 3) the generated weakly-supervised labels for the unannotated training medical data; and outputting the trained machine learning based task network.
1 . A computer-implemented method comprising:
receiving input medical data;
performing a medical analysis task using a trained machine learning based task network based on the input medical data; and
outputting results of the medical analysis task,
wherein the trained machine learning based task network is trained by:
receiving unannotated training medical data;
generating weakly-supervised labels for the unannotated training medical data using one or more trained machine learning based supervised learning networks;
training the machine learning based task network for performing the medical analysis task based on 1) the unannotated training medical data, 2) self-supervised labels for the unannotated training medical data learned via self-supervised learning, and 3) the generated weakly-supervised labels for the unannotated training medical data; and
outputting the trained machine learning based task network.
2 . The computer-implemented method of claim 1 , wherein training the machine learning based task network for performing the medical analysis task comprises:
assigning the generated weakly-supervised labels to features of a latent space of the machine learning based task network; and
training the machine learning based task network based on the assigned weakly-supervised labels.
3 . The computer-implemented method of claim 1 , wherein training the machine learning based task network for performing the medical analysis task comprises:
training the machine learning based task network using a cost function that incorporates the self-supervised labels and the generated weakly-supervised labels.
4 . The computer-implemented method of claim 1 , wherein training the machine learning based task network for performing the medical analysis task comprises:
fitting a probability distribution model to a latent space of the machine learning based task network to capture an uncertainty; and
training the machine learning based task network based on the uncertainty.
5 . The computer-implemented method of claim 1 , wherein generating weakly-supervised labels for the unannotated training medical data using one or more trained machine learning based supervised learning networks comprises:
generating initial labels using the one or more trained machine learning based supervised learning networks; and
filtering the initial labels using one or more out-of-domain probability distribution models to result in labels generated from input data that is in-domain of the one or more trained machine learning based supervised learning networks as the weakly-supervised labels.
6 . The computer-implemented method of claim 1 , wherein the machine learning based task network comprises an encoder and a plurality of decoders and training the machine learning based task network for performing the medical analysis task comprises:
training the machine learning based task network to perform a plurality of tasks performed respectively using one of the plurality of decoders.
7 . The computer-implemented method of claim 1 , wherein the machine learning based task network comprises an encoder and a plurality of decoders and training the machine learning based task network for performing the medical analysis task comprises:
training an initial decoder of the plurality of decoders to generate a reconstructed image based on features generated by the encoder; and
training one or more additional decoders of the plurality of decoders to respectively perform one or more tasks based on the reconstructed image.
8 . The computer-implemented method of claim 1 , wherein the machine learning based task network comprises an encoder and a plurality of decoders and training the machine learning based task network for performing the medical analysis task comprises:
finetuning the machine learning based task network using one or more of the plurality of decoders for domain adaptation.
9 . An apparatus comprising:
means for receiving input medical data;
means for performing a medical analysis task using a trained machine learning based task network based on the input medical data; and
means for outputting results of the medical analysis task,
wherein the trained machine learning based task network is trained by:
receiving unannotated training medical data;
generating weakly-supervised labels for the unannotated training medical data using one or more trained machine learning based supervised learning networks;
training the machine learning based task network for performing the medical analysis task based on 1) the unannotated training medical data, 2) self-supervised labels for the unannotated training medical data learned via self-supervised learning, and 3) the generated weakly-supervised labels for the unannotated training medical data; and
outputting the trained machine learning based task network.
10 . The apparatus of claim 9 , wherein training the machine learning based task network for performing the medical analysis task comprises:
assigning the generated weakly-supervised labels to features of a latent space of the machine learning based task network; and
training the machine learning based task network based on the assigned weakly-supervised labels.
11 . The apparatus of claim 9 , wherein training the machine learning based task network for performing the medical analysis task comprises:
training the machine learning based task network using a cost function that incorporates the self-supervised labels and the generated weakly-supervised labels.
12 . The apparatus of claim 9 , wherein training the machine learning based task network for performing the medical analysis task comprises:
fitting a probability distribution model to a latent space of the machine learning based task network to capture an uncertainty; and
training the machine learning based task network based on the uncertainty.
13 . The apparatus of claim 9 , wherein generating weakly-supervised labels for the unannotated training medical data using one or more trained machine learning based supervised learning networks comprises:
generating initial labels using the one or more trained machine learning based supervised learning networks; and
filtering the initial labels using one or more out-of-domain probability distribution models to result in labels generated from input data that is in-domain of the one or more trained machine learning based supervised learning networks as the weakly-supervised labels.
14 . A non-transitory computer readable medium storing computer program instructions, the computer program instructions when executed by a processor cause the processor to perform operations comprising:
receiving input medical data;
performing a medical analysis task using a trained machine learning based task network based on the input medical data; and
outputting results of the medical analysis task,
wherein the trained machine learning based task network is trained by:
receiving unannotated training medical data;
generating weakly-supervised labels for the unannotated training medical data using one or more trained machine learning based supervised learning networks;
training the machine learning based task network for performing the medical analysis task based on 1) the unannotated training medical data, 2) self-supervised labels for the unannotated training medical data learned via self-supervised learning, and 3) the generated weakly-supervised labels for the unannotated training medical data; and
outputting the trained machine learning based task network.
15 . The non-transitory computer readable medium of claim 14 , wherein training the machine learning based task network for performing the medical analysis task comprises:
assigning the generated weakly-supervised labels to features of a latent space of the machine learning based task network; and
training the machine learning based task network based on the assigned weakly-supervised labels.
16 . The non-transitory computer readable medium of claim 14 , wherein training the machine learning based task network for performing the medical analysis task comprises:
training the machine learning based task network using a cost function that incorporates the self-supervised labels and the generated weakly-supervised labels.
17 . The non-transitory computer readable medium of claim 14 , wherein the machine learning based task network comprises an encoder and a plurality of decoders and training the machine learning based task network for performing the medical analysis task comprises:
training the machine learning based task network to perform a plurality of tasks performed respectively using one of the plurality of decoders.
18 . The non-transitory computer readable medium of claim 14 , wherein the machine learning based task network comprises an encoder and a plurality of decoders and training the machine learning based task network for performing the medical analysis task comprises:
training an initial decoder of the plurality of decoders to generate a reconstructed image based on features generated by the encoder; and
training one or more additional decoders of the plurality of decoders to respectively perform one or more tasks based on the reconstructed image.
19 . The non-transitory computer readable medium of claim 14 , wherein the machine learning based task network comprises an encoder and a plurality of decoders and training the machine learning based task network for performing the medical analysis task comprises:
finetuning the machine learning based task network using one or more of the plurality of decoders for domain adaptation.
20 . A computer-implemented method comprising:
receiving unannotated training medical data;
generating weakly-supervised labels for the unannotated training medical data using one or more trained machine learning based supervised learning networks;
training a machine learning based task network for performing a medical analysis task based on 1) the unannotated training medical data, 2) self-supervised labels for the unannotated training medical data learned via self-supervised learning, and 3) the generated weakly-supervised labels for the unannotated training medical data; and
outputting the trained machine learning based task network.