Deep learning driven radiographic classification of primary bone tumors using attention augmented hybrid models

Accurate classification of primary bone tumors is necessary for timely diagnosis and effective treatment planning, particularly given the complex radiographic heterogeneity exhibited by tumor subtypes. The present study introduces two novel deep learning models, including a Convolutional Neural Network Transformer (CNNT) hybrid and a Residual Network 50 (ResNet50) model, augmented by a Convolutional Block Attention Module (CBAM) to enhance feature discrimination and contextual understanding in radiographic images. The models are trained and validated on the Bone Tumor X-ray Radiograph Dataset (BTXRD) dataset of 3,746 labeled radiographs containing nine tumor subtypes. To counter the effects of noise and class imbalance, advanced preprocessing methods like Block Matching 3D Filtering (BM3D) and data balancing using the Synthetic Minority Over sampling Technique (SMOTE) are employed. Extensive testing demonstrates that our approaches outperform current state of the art models, such as ResNet50, EfficientNet version B3 (EfficientNet-b3), You Only Look Once version 8 classification (YOLOv8s-cls), and Deep Supervision Network (DS-Net). Specifically, the ResNet50-CBAM architecture achieves an F1-score of 0.9759, an AUC-ROC score of 0.984, mean accuracy of CBAM 97.41% and a Cohen's Kappa score of 0.9718, outperforming existing benchmarks for binary tumor classification. The CNNT model also achieves competitive performance, reaching an F1-score of 0.9595 with an accuracy of 92.56%. Incorporating attention mechanisms and dataset guided preprocessing renders this framework appropriate for practical clinical settings. The findings of this research have significant implications for the healthcare sector by introducing a scalable, interpretable, and highly accurate Artificial Intelligence (AI) based diagnostic system that can support radiologists in timely diagnoses and decision making processes, ultimately contributing to better patient outcomes and alleviating the diagnostic burden in musculoskeletal oncology.