AI bone lesion classifier with sensitivity-driven optimization for radiographs

Aim: This study aims to develop a deep learning classifier for detecting primary bone lesions on radiographs, emphasizing high sensitivity while maintaining practical clinical usability. Material and Methods: Radiographs of the upper and lower extremities were reviewed by board-certified radiologists and categorized into two groups: “Normal” (without bone lesions) and “Abnormal” (with bone lesions). The final dataset comprised 1,177 radiographs from 310 patients, including 547 abnormal and 630 normal cases. The MobileNetV2 architecture was trained with a sensitivity-driven approach designed to minimize false negatives. Model performance was evaluated on a hold-out test set, and attention maps were generated to enhance interpretability and visualize regions contributing to the model's decisions. Results: The model was tested on a naïve hold-out test set. The results received on the test set: sensitivity of 96.6%, specificity of 82.2%, accuracy of 87.9%, area under the curve (AUC) of 0.94, and 95% confidence interval of [0.901, 0.981]. Conclusion: The study demonstrates the feasibility of deploying AI-based tools for radiographic detection of bone tumors with a sensitivity-focused optimization. These tools have the potential to enhance diagnostic accuracy, reduce diagnostic delays, and support population health initiatives.