Deep Learning-Based Classification of Primary Bone Tumors on Radiographs: A Preliminary Study

Background: To develop a deep learning model to classify primary bone tumors from preoperative radiographs and compare performance with expert radiologists. Methods: A total of 1,057 patients (2,233 images) with histologically confirmed primary bone tumors and pre-operative radiographs were identified from three institutions’ pathology databases. Lesions were manually segmented by a radiologist. Binary discriminatory capacity (benign versus not-benign and malignant versus not-malignant) and three-way classification (benign versus intermediate versus malignant) performance of our model were evaluated. Final model performance was compared with two experts’ interpretation. Findings: For benign vs. not benign, the model achieved 82.3% accuracy (AUC [area under curve]: 0.895) compared to 84.1% and 81.0% for experts 1 and 2 (p=0.25 for expert 1 and p=0.34 for expert 2). For malignant vs. not malignant, the model achieved 85.6% accuracy (AUC: 0.908) vs. 84.9% and 85.5% for experts 1 and 2 (p=0.79 for expert 1 and p=0.96 for expert 2). For three-way classification, the model achieved 75.0% accuracy vs. 74.7% and 72.0% for experts 1 and 2 (p=0.76 for expert 1 and p=0.17 for expert 2). Interpretation: Deep learning can classify primary bone tumors using conventional radiographs in a multi-institutional dataset with similar accuracy compared to experts. Future study will focus on development of a fully automatic pipeline including lesion localization and further validation on additional datasets. Funding Statement: This project was supported by the National Institutes of Health under Award Number R03CA235202 and National Natural Science Foundation of China grant under Award Number 8181101287 to Harrison X. Bai. Declaration of Interests: No potential conflicts of interest were disclosed. Ethics Approval Statement: The study was conducted in accordance with Declaration of Helsinki and approved by the Institutional Review Boards at all three institutions.