Domain adaptive detection framework for multi-center bone tumor detection on radiographs

Abstract

Automatic bone tumor detection on radiographs is crucial for reducing mortality from bone cancer. However, the performance of the detection methods may be considerably affected when deployed to bone tumor data in a distinct domain, which could be attributed to the differences in the imaging process and can be solved by training with a large amount of annotated data. However, these data are difficult to obtain in clinical practice. To address this challenge, we propose a domain-adaptive (DA) detection framework to effectively bridge the domain gap of bone tumor radiographs across centers, consisting of four parts: a multilevel feature alignment module (MFAM) for image-level alignment, Wasserstein distance critic (WDC) for quantization of feature distance, instance feature alignment module (IFAM) for instance-level alignment, and consistency regularization module (CRM), which maintains the consistency between the domain predictions of MFAM and IFAM. The experimental results indicated that our framework can improve average precision (AP) with an intersection over union threshold of 0.2 (AP@20) on the source and target domain test sets by 1 % and 8.9 %, respectively. Moreover, we designed a domain discriminator with an attention mechanism to improve the efficiency and performance of the domain-adaptative bone tumor detection model, which further improved the AP@20 on the source and target domain test sets by 2 % and 10.7 %, respectively. The proposed DA model is expected to bridge the domain gap and address the generalization problem across multiple centers.