Medical Transformer: Gated Axial-Attention for Medical Image Segmentation

Over the past decade, deep convolutional neural networks have been widely adopted for medical image segmentation and shown to achieve adequate performance. However, due inherent inductive biases present in architectures, they lack understanding of long-range dependencies image. Recently proposed transformer-based architectures that leverage self-attention mechanism encode learn representations are highly expressive. This motivates us explore solutions study feasibility using network tasks. Majority existing vision applications require large-scale datasets train properly. compared applications, imaging number data samples is relatively low, making it difficult efficiently transformers applications. To this end, we propose a gated axial-attention model which extends by introducing an additional control module. Furthermore, effectively on images, Local-Global training strategy (LoGo) further improves Specifically, operate whole patches global local features, respectively. The Medical Transformer (MedT) evaluated three different achieves better performance than other related architectures. Code: https://github.com/jeya-maria-jose/Medical-Transformer