Multi-scale attention U-net for segmenting clinical target volume in graves’ ophthalmopathy

Graves’ ophthalmopathy (GO) is an autoimmune inflammatory disorder associated with thyroid disease, and radiotherapy is an effective treatment that causes few side effects among patients in the moderate-to-severe stage. The clinical target volume (CTV) refers to tissues with potential tumor spread or subclinical diseases, where accurate segmentation of these tissues is required for the successful radiotherapy. Traditional segmentation methods for the CTV are based on low-level hand-crafted features that require significant domain knowledge and sensitive to the variations. To overcome these shortcomings, a novel neural network architecture called multi-scale attention U-Net (MAU-Net) is proposed to automatically segment the CTV by computed tomography for GO disease. Abstract features ranging from low- to high-levels are extracted by a deep residual network, then processed by a multi-scale module composed of multiple convolutional operations to accommodate various scales of the CTV. A novel attention module is proposed and applied following the multi-scale module, which uses signals from high-level features to selectively highlight the low-features. A total of 178 CT cases are used to train and evaluate the proposed MAU-Net. The experimental results show that MAU-Net achieves higher segmentation accuracy than the state-of-the-art methods. The MAU-Net converges more quickly in the training phase, and achieves lower error on the validation dataset than the vanilla U-Net.