U-shaped Transformers for 3D Lung Cancer Segmentation

Abstract

3-Dimensional (3D) image segmentation in medical images is essential for early detection and diagnosis of diseases. It also aids in effective monitoring and treatment preparation. Traditional methods of delineating the image manually requires anatomical knowledge and is error-prone, cumbersome and expensive. Deep learning methods, especially V-shaped Convolutional Neural Network (CNN) architectures have achieved state-of-the-art performance on 2-Dimensional (2D) clinical image data. However, when it comes to 3D medical images, they suffer from anisotropy which is non-homogeneity in all directions. This paper shows that conventional convolution-based networks are insufficient to accurately segment this kind of data and proposes a ‘U-shaped’ transformer-based network, leveraging the self-attention mechanism to achieve better segmentation results. The proposed model outperforms baseline convolution-based models in 3D lung cancer segmentation.