Research on few-shot bone tumor classification method based on C-MSPNet

Abstract

Aiming at the insufficient utilization of clinical information, limited adaptability to lesion scales, and lack of clinical-prior guidance in feature learning of traditional prototypical network models for bone tumor image classification, this paper proposes a Clinical-guided Multi-Scale Prototypical Network (C-MSPNet). The method jointly models local, intermediate, and global image features and incorporates clinical information, including age, gender, lesion site, and joint involvement, into the prototypical learning process as auxiliary conditions. Under a unified few-shot evaluation protocol on the BTXRD dataset with 100 fixed test episodes, C-MSPNet achieved a Macro-F1 of 87.04% ± 6.18%, an AUC of 92.43%, a malignant tumor sensitivity of 84.33%, and a benign tumor specificity of 90.00%, with a 95% confidence interval of 85.82%–88.25% for Macro-F1. Compared with the evaluated few-shot learning methods, C-MSPNet improved Macro-F1 by 5.27 percentage points over the best-performing baseline MatchingNet and improved AUC and malignant tumor sensitivity by 3.46 and 9.40 percentage points, respectively, over PrototypicalNet, demonstrating better overall classification performance and malignant case detection. The results suggest that, under the current retrospective dataset and experimental protocol, jointly modeling multi-scale image information and clinical conditions has the potential to improve benign–malignant bone tumor classification. Because the public dataset lacks patient-level identifiers and no independent external validation has been conducted, the reported results may be affected by potential data leakage and selection bias and cannot be used to infer cross-center generalizability or clinical applicability.