Hierarchical visual attention and stage-wise attribution interpretation for heel spur classification in radiographic images

Heel spurs are bony outgrowths associated with chronic heel pain and plantar fasciitis. Automated detection from X-ray images remains challenging due to low contrast and anatomical variability. This work presents an attention-augmented deep learning framework for heel spur classification, integrating a Convolutional Block Attention Module (CBAM) and Particle Swarm Optimisation (PSO). A gradient-based stage-wise attribution analysis is proposed to quantify hierarchical feature importance, revealing that high-level semantic stages (stage_3_downsample) dominate prediction decisions. The model achieves 96.97 % accuracy and 0.993 macro-AUC on a clinical dataset. Grad-CAM visualisation and OCR-enhanced metadata extraction improve interpretability and clinical applicability. This study provides an interpretable solution for automated heel spur diagnosis. The codes that are used in this study are available at: https://github.com/raghavbadri06-gif/Heelspur/tree/main.