A Novel Self-Supervised Accurate Biomedical Image Segmentation Scheme Using Hybrid Learning Enabled Classification Logic

Correct biomedical image segmentation is an important problem because of the lack of annotated data and because tissue boundaries of multi-modal clinical scans are complicated. The current paper will present a new self-supervised biomedical image segmentation architecture that combines hybrid learning based on convolutional modules and transformer-based modules along with a classification-driven feedback system. With the help of contrastive and predictive self-supervised tasks the encoder learns strong representations without having to be manually annotated. A dual-branch structure is a combination of local texture learning (ConvNeXt) and global context modeling (Swin Transformer), and then a U-Net architecture is used with an improved decoder using attention-based skip connections. A parallel branch of a classifier is used as a way of semantic feedback to tune the segmentation map by using class-aware attention gating. Experimental analyses on BRaTS, ISIC and MoNuSeg datasets show better results with Dice score of 94.6, precision of 94.2 and recall of 93.9 and F1-score of 94.0. The proposed model, also, is the best, when working with the state-of-the-art baselines, the maximum label scarcity, and 82.7% Dice score is reached with 5% labeled data only. These findings confirm the usefulness of the framework in low-annotation and cross-domain. A combination of self-supervised representation learning and classification-conscious segmentation is a promising field of scalable, precise, and generalizable biomedical image analysis.