A novel consistency-guided and pseudo labeled training for weakly-semi-supervised medical image segmentation

Medical image segmentation is crucial for disease diagnosis but is often hindered by the scarcity of annotated data, which is time-consuming and expensive to acquire. To address this challenge, this paper proposes a novel semi-supervised learning framework for medical image segmentation that integrates consistency regularization and pseudo-labeling techniques. The framework utilizes an ensemble of four models with diverse architecture (YOLOv11 and Vision Transformers) to enhance model diversity and generalization. Two models are trained in a supervised manner on limited labeled data, while two auxiliary models are trained using weakly supervised learning. A Weak Pseudo Labeling (WPL) strategy employing superpixels generates initial labels for unlabeled data used by the auxiliary models. Data augmentation, including the Mix-up technique, is applied to enhance robustness, particularly for the auxiliary models trained on noisy, combined inputs. A novel Segmentation Attention-based Head (SAH) integrating Spatial and Channel Attention mechanisms is introduced to refine feature maps and improve segmentation accuracy. Furthermore, a Final Pseudo Labeling (FPL) process refines pseudo-labels by selecting high-confidence predictions based on inter-model agreement, iteratively expanding the labeled set. Extensive experiments on skin lesion datasets demonstrate that the proposed method outperforms state-of-the-art semi-supervised techniques. Using only 100 labeled samples, it achieved high Jaccard Similarity Index (JSI) scores of 86.16% on ISIC2018 and 83.91% on ISIC2017. Robustness evaluations on the unseen PH₂ dataset further confirmed the framework's generalization, yielding a JSI of 86.46%. The proposed approach offers a highly effective and data-efficient solution for semi-supervised medical image segmentation.