PERFORMANCE ENHANCEMENT OF SEGMENTATION ALGORITHMS IN MEDICAL IMAGING USING TRANSFORMER-BASED ARCHITECTURES

Abstract

Medical image segmentation is a foundational task in clinical diagnosis, surgical planning, and treatment monitoring, requiring precise delineation of anatomical structures and pathological regions from complex imaging modalities. Conventional convolutional neural network (CNN)-based segmentation frameworks, including the widely adopted U-Net architecture, have demonstrated substantial success but remain constrained by their inherently local receptive fields, limiting their capacity to model long-range spatial dependencies essential for accurate delineation of irregular and heterogeneous lesions. Transformer-based architectures, originally developed for natural language processing, have recently emerged as powerful alternatives capable of capturing global contextual relationships through self-attention mechanisms. This paper presents a comprehensive review and experimental analysis of transformer-based segmentation models — including TransUNet, Swin-UNet, nnFormer, and MedT — evaluated across multiple medical imaging modalities: MRI, CT, and histopathology. Experiments conducted on the Synapse multi-organ segmentation dataset and ACDC cardiac segmentation benchmark demonstrate that transformer-based hybrid architectures consistently outperform CNN baselines, achieving mean Dice Similarity Coefficients (DSC) of up to 79.1% and 90.4% respectively, while reducing Hausdorff Distance (HD95) by 12–18%. The study further analyses the role of patch embedding, multi-scale feature fusion, and positional encoding strategies in determining segmentation accuracy and computational efficiency. Findings confirm that transformers with hierarchical attention mechanisms offer the most favourable trade-off between global context modelling and computational cost, establishing a clear pathway for clinical deployment of transformer-based segmentation systems.