Active learning for improved semi-supervised semantic segmentation in echocardiography

Accurate left ventricle segmentation in echocardiography is essential for assessing cardiac health. Traditional medical image segmentation often requires large, labeled datasets, which are challenging and costly to obtain. This study introduces a hybrid methodology that combines semi-supervised learning (SSL) and active learning (AL) to reduce reliance on extensive annotations while ensuring high segmentation accuracy. Our primary objective is to enhance U-Net performance by leveraging a limited set of labeled data alongside a substantial pool of unlabeled data. We selectively label the most informative samples through targeted active learning strategies. Segmentation effectiveness is evaluated using metrics such as the Dice coefficient and Hausdorff distance. We train the U-Net model on both labeled and unlabeled datasets through techniques like pseudo-labeling and consistency regularization, implementing uncertainty-based active learning to annotate the most uncertain samples, thus streamlining the labeling process. We anticipate this approach will reduce the volume of labeled data needed by 70-80%, delivering state-of-the-art segmentation performance in clinical settings with minimal annotation costs. Ultimately, our hybrid models aim to surpass traditional fully-supervised models in accuracy and efficiency.