Efficient Utilization of Unlabeled Data during Training of Semi-Supervised Learning Model

Motivation: Current simultaneous semi-supervised training method uses an equal number of labeled and unlabeled data samples (balanced sampling) in each epoch, which seems to be an inefficient utilization of all the unlabelled data. Goal(s): To develop a training approach that efficiently utilizes all the unlabeled dataset in each training epoch. Approach: We propose a training method that first uses a balanced sampling to train the model, followed by training using unbalanced sampling. Results: We report an improvement in the segmentation performance with the proposed training approach against conventional training approach (Dice Score: 0.86 ± 0.02 versus 0.84 ± 0.04, p<0.001). Impact: We propose a simultaneous semi-supervised model training methodology that ensures efficient utilization of all the available unlabeled data in each training epoch.