An attention-enhanced deep learning framework for robust bone tumor screening from radiographs

Timely diagnosis and efficient treatment planning depend on the early and precise detection of bone cancers from X-ray imaging.Using the publicly accessible Bone Tumor X-ray Radiograph Dataset (BTXRD), this study assesses deep learning models for both binary (normal vs. tumor) and multiclass (normal, benign, malignant) classification: MobileNetV2, EfficientNetB0, and a hybrid ResNet50 + Transformer. All models reached stable convergence for binary classification, however the hybrid ResNet50+Transformer model outperformed the others (AUC = 0.80, accuracy = 71%). After targeted augmentation and class weighting, the same model demonstrated higher discrimination (macro-AUC = 0.91) in the multiclass test, especially enhancing memory for malignant cases. The results demonstrate that complex bone tumor patterns are better represented when convolutional feature extraction and transformer based attention are combined. This method shows how hybrid CNN–Transformer frameworks can be used for accurate and effective radiograph based bone tumor screening.