XChondNet: Explainable Chondrogenic Tumor Diagnosis by Spatial-Context-Aware Synergistic Deep Feature Fusion

As a common benign bone tumor, the pathological diagnosis of chondrogenic tumors needs to accurately analyze the calcification pattern of the cartilage matrix, key spatial topology information, and other important indicators. However, chondrogenic tumors are a heterogeneous group of tumors and a rare disease: doctors lack sufficient reference data and experience in diagnosis. At the same time, complex spatial structural information also increases the difficulty of diagnosis, which leads to inconsistencies in the diagnosis's results of different doctors. In contrast, with the development of technology, artificial intelligence (AI), with its fast, accurate, and robust characteristics, can effectively improve the efficiency and accuracy of diagnosis. However, the application of AI in this field is still unrecognized, so it is urgent to develop a model that can help doctors in diagnosis to improve accuracy and efficiency. In this study, we propose XChondNet, an explainable spatial-context-aware synergistic deep feature fusion model for WSI-based chondrogenic tumor classification. The model proposes a fusion mechanism of pathological and positional features so that the model can effectively perceive spatial structural information and a parallel classifier mechanism based on potential coding, which can effectively solve the problem of class imbalance in chondrogenic tumor data. We evaluated the XChondNet model on our chondrogenic tumor dataset and the experimental results verified its effectiveness in the classification of the chondrogenic tumor subtype. In the test phase, XChondNet consistently achieved superior performance across different feature extractors. Compared with two strong MIL baselines, DTFD-MIL and RRT-MIL, XChondNet improved the average ACC from 85.16% to 86.58% (1.42 percentage points), with statistically significant differences confirmed by a paired t-test (p=2.61×10-5) and Wilcoxon signed-rank test (p=0.0078). Moreover, the attention maps of XChondNet explicitly reflect regions consistent with pathologists' diagnostic concerns, enhancing the interpretability of AI-assisted diagnosis.