Explainable cross-anatomical bone fracture classification using spatial-channel attention and CNN-LSTM-based feature interdependence

Precise detection of bone fractures across various anatomical regions using X-ray imaging is essential for orthopedic diagnosis and treatment. However, the artifacts, varying contrast, overlapping structure, and lack of experienced surgeons make it challenging. Although conventional CNNs can classify fractured images well, they often struggle with limited, complex, and diverse fracture datasets, including various types of bone fractures, leading to inconsistent outcomes. This paper presents a hybrid model comprising convolutional layers, a modified channel- and spatial-attention mechanism, and LSTM layers to improve cross-anatomical fracture detection. The modified attention layer enhances the features of each channel, focusing on specific regions within the spatial features produced by convolutional layers. The LSTM layers encapsulate feature interdependence, enabling the network to assimilate prior spatial characteristics. The proposed network is tested on two datasets containing bone fractures from various body regions. The proposed model is analyzed using diagnostic parameters, including sensitivity, specificity, and likelihood ratio. Subsequently, an ablation study is presented, comparing it to existing literature using precision, recall, and F1 score. The results show that the model succeeded with a 98% F1-score on both datasets. A negative likelihood ratio below 0.1, with a sensitivity of about 97.04%, indicates the model’s robustness on a diverse dataset. Furthermore, t-SNE analysis, Silhouette score, and Grad-CAM heatmaps are proposed to visualize and understand how features propagate across layers and contribute to the decision-making process for classifying bone fractures. Experimental findings demonstrate that the proposed method outperforms current techniques, underscoring its applicability to clinical radiographic analysis.