Figure 4.

Study flow of DNN learning and evaluation. (a) MRI slices of uterine leiomyomas and sarcomas are augmented to 875,000 slices. In one epoch, 35,000 slices are selected randomly out of 875,000 slices and the model repeats learning 50 times. The ratio of the learning set to the evaluation set is 5:1, which is cross validated. The DNN models are evaluated as “a uterine sarcoma” or “a uterine leiomyoma” using either a single-model prediction or ensemble prediction. The augmentation method we adopted was a very general approach that included flips, rotations, zooms, and changes to the brightness. (b) For evaluation, 24 sets of ensemble predictions are performed along with single-model predictions. The predictions of the ensemble model combine the results of 23 models.