Self-Supervised Learning for Tumor Segmentation in MRI Scans

Precise tumor segmentation from MRI scans is an important medical image analysis task, but is generally marred by a lack of annotated datasets to train supervised deep learning algorithms on. To overcome this, the present paper introduces a self-supervised learning (SSL) approach that utilises massive sets of unlabelled MRI scans to pretrain models and then fine-tunes them using small collections of labelled examples. The strategy adopts masked image modeling and contrastive learning as pretext tasks and fine-tunes through a U-Net-based segmentation network. The model is capable of learning spatial and contextual information for tumor detection with this strategy. Experimental results indicate that SSL + Fine-Tuned U-Net performs better than the classical supervised models with a Dice Similarity coefficient of 89.6% and an IoU of 81.4%, and its precision and recall are enhanced. These results suggest that SSL is capable of greatly improving segmentation accuracy and robustness, even from limited amounts of labeled data. This work is a contribution as a scalable and effective methodology towards improving AI-assisted diagnosis in resource-limited clinical settings.