Spinal curve assessment of idiopathic scoliosis with a small dataset via a multi-scale keypoint estimation approach

Idiopathic scoliosis (IS) is the most common type of spinal deformity, which leads to severe pain and potential heart and lung damage. The clinical diagnosis and treatment strategies for IS highly depend on the radiographic assessment of spinal curve. With improvements in image recognition via deep learning, learning-based methods can be applied to facilitate clinical decision-making. However, these methods usually require sufficiently large training datasets with precise annotation, which are very laborious and time-consuming especially for medical images. Moreover, the medical images of serious IS always contain the blurry and occlusive parts, which would make the strict annotation of the spinal curve more difficult. To address these challenges, we utilize the dot annotations approach to simply annotate the medical images instead of precise annotation. Then, we design a multi-scale keypoint estimation approach that incorporates Squeeze-and-Excitation(SE) blocks to improve the representational capacity of the model, achieving the assessment of spinal curve without large-size dataset. The proposed approach uses pose estimation framework to detect keypoints of spine with simple annotation and small-size dataset for the first time. Finally, we conduct experiments on a collected clinical dataset, and results illustrate that our approach outperforms the mainstream approaches.