Cross modality training of machine learning models

1. A computer implemented method for training a visual machine learning (ML) model component and a natural language processing (NLP) ML model component, comprising:

providing a training dataset including, for each of a plurality of sample individuals, a medical image and a corresponding text based report;

providing the NLP ML model component for generating an outcome of at least one NLP category in response to an input of a target text based report;

providing the visual ML model component for generating an outcome of at least one visual finding in response to an input of a target image; and

concurrently training the NLP ML model component and the visual ML model component using the training dataset, by:

training the NLP ML model using an input of the text based reports of the training dataset and a ground truth comprising the outcome of the at least one visual finding generated by the visual ML model in response to an input of the images corresponding to the text based reports of the training dataset;

training the visual ML model using an input of the images of the training dataset and a ground truth comprising the outcome of the at least one NLP category generated by the NLP ML model in response to an input of the text based reports corresponding to the images of the training dataset;

computing a correlation value indicative of a correlation between the at least one NLP category outcome of the NLP ML model and the at least one visual finding outcome of the visual ML model for an input of an image and corresponding text based report; and

in response to the correlation value being below a threshold indicative of dis-correlation between the at least one NLP category outcome of the NLP ML model and the at least one visual finding outcome of the visual ML model, storing the image and corresponding text based report in a user-training dataset; and

providing the user-training dataset for presentation on a display.

2. The method of claim 1 , wherein the NLP ML model is trained using a supervised approach with the input of the based reports and the ground truth outcome of the visual ML model, and concurrently the visual ML model is trained using a supervised approach with the input of the images and the ground truth outcome of the NLP ML model.

3. The method of claim 1 , wherein the concurrently training is performed iteratively.

4. The method of claim 1 , further comprising:

prior to the concurrently training, weakly labelling a subset of the text based reports of the training dataset with a weak label indicative of presence or absence of the at least one NLP category in respective target based reports; and

wherein the concurrently training is performed using the training dataset with weak labels of the text based reports.

5. The method of claim 4 , wherein weakly labelling comprises weakly labelling about 5-20% of the text based reports of the training dataset with the weak label.

6. The method of claim 4 , wherein weakly labelling comprises automatically weakly labelling the subset of the text based reports using a simple set of rules.

7. The method of claim 1 , wherein the at least one NLP category outcome of the NLP ML model and the at least one visual finding outcome of the visual ML model are from a common set and of a same format.

8. The method of claim 1 , wherein each of the at least one NLP category outcome of the NLP ML model and the at least one visual finding outcome of the visual ML model is a binary classification indicative of positive or negative finding found in the image and corresponding text based report.

9. The method of claim 1 , wherein the at least one NLP category outcome of the NLP ML model component is an indication of a visual finding depicted in an image corresponding to a text based report inputted into the NLP ML model component, and the at least one visual finding outcome of the visual ML model component is an indication of the visual finding depicted in the image corresponding to the text based report inputted into the NLP ML model component.

10. The method of claim 1 , wherein concurrently training comprises concurrently training the NLP ML model component and the visual ML model component using a combined visual and NLP consensus loss function.

11. The method of claim 1 , wherein the combined visual and NLP consensus loss function comprises a cross model consensus loss function that encourages high consensus between the NLP ML model and the visual ML model.

12. The method of claim 10 , wherein concurrently training further comprises training the NLP ML model component using an NLP loss function that is computed for the training of the NLP ML model and excludes data obtained from the training of the visual ML model component.

13. The method of claim 12 , wherein the NLP ML model comprises a binary classifier and the NLP loss function comprises a standard binary cross entropy loss.

14. The method of claim 12 , wherein the NLP loss function penalizes the NLP ML model for errors made during an initial inaccurate labeling of a subset of text based reports of the training dataset made prior to the concurrently training.

15. The method of claim 1 , wherein the visual ML model component is implemented as a neural network.

16. The method of claim 1 , wherein the NLP ML model component is implemented as a neural network.

17. The method of claim 1 , wherein a target text report is inputted into an NLP processing path comprising the NLP ML model component that generates the NLP category, and a target image corresponding to the target text report is inputted in a visual processing path comprising the visual ML model component that generates the at least one visual finding, wherein the NLP processing path and the visual processing path are concurrently executed during the concurrent training.