Anatomical Landmark Detection Using a Feature-Sharing Knowledge Distillation-Based Neural Network

Abstract

Existing anatomical landmark detection methods consider the performance gains under heavyweight network architectures, which lead to models tending to have poor scalability and cost-effectiveness. To solve this problem, state-of-the-art knowledge distillation (KD) methods are proposed. However, they only require the teacher model to guide the output of the final layer of the student model. In this way, the semantic information learned by the student model is very limited. Different from previous works, we propose a novel KD-based model-training strategy, named feature-sharing fast landmark detection (FSF-LD), which focuses on intermediate features and effectively transfers richer spatial information from the teacher model to the student model. Moreover, to generate richer and more reliable knowledge, we propose a multi-task learning structure to pretrain the teacher model before FSF-LD. Finally, a tiny and effective anatomical landmark detection model is obtained. We evaluate our proposed FSF-LD on a public 2D hand radiograph dataset, a public 2D cephalometric radiograph dataset and a private 2D hip radiograph dataset. On the 2D hand dataset, our FSF-LD has 11.7%, 12.1%, 12.0,% and 11.4% improvement on SDR (r = 2 mm, r = 2.5 mm, r = 3 mm, r = 4 mm) compared with other KD methods. The results suggest the superiority of FSF-LD in terms of model performance and cost-effectiveness. However, it is a challenge to further improve the detection accuracy of anatomical landmarks and realize the clinical application of the research results, which is also our next plan.