Active Learning and Adaptive Semi-Supervised Approach for Medical Image Segmentation

Abstract

Supervised Deep Learning approaches have proven highly effective in analyzing and segmenting medical images. However, their reliance on labeled data limits their scalability and generalization across various medical pathologies. Attaining reliable and well-aligned predictions requires ongoing research in order to develop alternative approaches, such as semi-supervised learning. Semi-supervised learning leverages the abundance of the vast amount of unlabeled data along with the small labeled dataset to enhance the model’s performance, thus reducing the intensive effort of manual labeling. However, generating pseudo-labels for extensive datasets is computationally expensive. Therefore, this approach is combined with Active Learning. Accordingly, in this study, we proposed an end-to-end automated framework that selects the most informative images for further pseudo-labeling validation using a geometrical method. Our proposed approach demonstrated its effectiveness in reducing manual pseudo-labeling while achieving a higher Dice Coefficient of $99.59 \%$, leading to accurate clinical decisions.