A Calibrated Deep Learning Framework Integrating Spatial Annotations and Clinical Metadata for Safe Three-Class Bone...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Forensic Anthropology

Background:

• Accurate bone lesion classification is vital for clinical decisions and forensic identification.
• Current deep learning models often overlook spatial annotations and clinical metadata in radiographic analysis.
• There is a need for advanced AI frameworks that incorporate these crucial data points.

Purpose of the Study:

• To develop a region-of-interest (ROI)-guided deep learning framework for three-class bone lesion classification (Normal, Benign, Malignant).
• To integrate clinical metadata into the deep learning model for enhanced classification accuracy.
• To rigorously assess the clinical safety profile of the developed framework.

Main Methods:

• Utilized the BTXRD dataset comprising 3746 radiographs (Normal, Benign, Malignant).
• Employed an EfficientNetV2-S backbone with an 11-dimensional metadata Multi-Layer Perceptron (MLP) trained on ROI-cropped regions.
• Implemented advanced training techniques including Focal Loss, Mixup/CutMix augmentations, Stochastic Weight Averaging, and Test-Time Augmentation.

Main Results:

• Achieved high performance metrics: 96.05% accuracy, 93.94% balanced accuracy, 92.62% macro F1-score, and 99.21% macro-AUC.
• Demonstrated a clinically safe error pattern with near-zero Malignant-to-Normal misclassifications (0.29%).
• The minority Malignant class achieved an F1-score of 83.53% despite its low representation (9.1% of the dataset).

Conclusions:

• ROI-guided deep learning with metadata fusion represents a state-of-the-art approach for bone lesion classification.
• The framework exhibits clinically safe error patterns and well-calibrated probability outputs.
• The model shows potential as a decision support tool in diagnostic radiology and forensic anthropology, pending external validation.