| Abstract [eng] |
Semi-supervised learning (SSL) is a promising approach to address the challenges of limited labeled data in medical image segmentation, particularly in 3D magnetic resonance imaging (MRI). The goal of semi-supervised learning is to learn patterns from unlabeled data, improving the accuracy of models trained on limited labeled datasets. However, SSL is an emerging field with numerous techniques being introduced, leading to ambiguity in method classification, and only a limited number of comprehensive reviews on semi-supervised learning are available. This research systematically investigates various semi-supervised learning techniques, focusing on their application to segmentation tasks involving the Left Atrium and BraTS-Africa datasets. A comprehensive literature review was conducted to identify and classify prominent semi-supervised methods, namely consistency regularization, pseudo-labeling, co-training, contrastive learning, adversarial learning, and hybrid methods. Comparative experiments were conducted, with techniques including Mean Teacher (MT), Deep Adversarial Networks (DAN), Adversarial Entropy Minimization (ADVENT), Cross Pseudo Supervision (CPS), Deep Co-Training (DCT), and Semi-Supervised Contrastive Consistency (SCC). These methods were evaluated using metrics such as Dice coefficient, Jaccard index, HD95, and ASD. Comparative experiments demonstrated that Cross Pseudo Supervision and Deep Co-Training outperformed other semi-supervised approaches, achieving results closer to fully supervised models, especially when applied to datasets with simpler structures, such as the Left Atrium. In Left Atrium case, contrastive learning (SCC) approach yielded the best scores, however, this approach did not work in BraTS-Africa case. Increasing the proportion of labeled data from 10% to 20% led to substantial improvements in segmentation performance, highlighting the importance of labeled data for model training. However, more complex datasets, like BraTS-Africa, posed additional challenges due to heterogeneous tumor regions, resulting in lower accuracy and precision as compared with Left Atrium dataset. |
