Abstract
The increasing demand for precise and scalable analysis of high-resolution medical images has exposed fundamental limitations in conventional convolutional neural network (CNN)-based segmentation models, particularly in capturing long-range spatial dependencies and contextual relationships. Vision Transformer (ViT) architectures have emerged as a promising alternative by leveraging self-attention mechanisms to model global context, thereby enhancing feature representation for complex medical imaging tasks. This study presents a comprehensive investigation of Vision Transformer architectures for high-resolution medical image segmentation and disease localization, integrating insights from foundational transformer models, hybrid CNN–transformer frameworks, and pure transformer-based segmentation approaches. A unified conceptual framework is proposed that combines patch-based tokenization, hierarchical transformer encoding, multi-scale feature fusion, and decoder-based reconstruction for dense segmentation, alongside an auxiliary localization module informed by attention and gradient-based interpretability techniques. The study systematically evaluates representative architectures, including TransUNet, UNETR, Swin-Unet, nnFormer, and hybrid fusion models, across widely used benchmark datasets such as BraTS, LIDC-IDRI, ISIC, and KiTS19. Performance is assessed using standard metrics including Dice coefficient, Intersection over Union (IoU), precision, recall, and localization accuracy. The analysis demonstrates that transformer-based and hybrid architectures consistently outperform traditional CNN models in high-resolution segmentation tasks, particularly in scenarios requiring global contextual understanding and fine-grained boundary delineation. Furthermore, the integration of attention-driven localization mechanisms enhances interpretability and supports clinically meaningful disease region identification. However, challenges related to computational complexity, data dependency, and deployment constraints remain significant. This work contributes a structured synthesis of current advancements, a conceptual design framework for transformer-based medical segmentation systems, and a comparative evaluation that highlights both the strengths and limitations of existing approaches. The findings provide a foundation for future research aimed at developing efficient, interpretable, and clinically deployable transformer-based solutions for medical image analysis.