A Simple Localization and Classification Framework based on Template Matching and Self-Attention for TAO treatment diagnosis

Treatment decision of thyroid-associated orbitopathy is crucial to alleviate different side effect of the disease. Moreover, recent studies show the possibility of neural network-based treatment prediction. However, existing neural networks are affected by regions with fewer disease symptoms. Moreover, fine-grained segmentation or cropping is time-consuming and requires a large resource of professionals. In this paper, a simple localization and classification framework is proposed based on template matching and self-attention. Our experimental results show that the proposed method performs statistically better than the existing method with the simple effort of pre-defining eight orbit templates.