UTNet: A Hybrid Transformer Architecture for Medical Image Segmentation

Transformer architecture has emerged to be successful in a number of natural language processing tasks. However, its applications medical vision remain largely unexplored. In this study, we present UTNet, simple yet powerful hybrid that integrates self-attention into convolutional neural network for enhancing image segmentation. UTNet applies modules both encoder and decoder capturing long-range dependency at different scales with minimal overhead. To end, propose an efficient mechanism along relative position encoding reduces the complexity operation significantly from \(O(n^2)\) approximate O(n). A new is also proposed recover fine-grained details skipped connections encoder. Our approach addresses dilemma requires huge amounts data learn inductive bias. layer design allows initialization networks without need pre-training. We have evaluated on multi-label, multi-vendor cardiac magnetic resonance imaging cohort. demonstrates superior segmentation performance robustness against state-of-the-art approaches, holding promise generalize well other segmentations.