Toward Automated Foot Type Diagnosis: A Deep Learning Approach with Attention Mechanism and Grad-CAM Interpretability

Plantar pressure distribution plays a critical role in assessing foot function, diagnosing foot disorders, and designing orthotics. Traditional machine learning methods have limitations in capturing complex spatial patterns and achieving high accuracy due to reliance on manual feature selection. These methods often rely on manually selected features and struggle to extract intricate relationships within the plantar pressure data, leading to suboptimal performance. Recently, hierarchical features can be automatically observed from raw data by mechine learning approaches. These methods can capture both local and global patterns, demonstra superior performance in extracting intricate features and delivery robust classification results. Nevertheless, their inherent complexity often leads to a lack of interpretability. In this paper, an application of deep learning method for plantar pressure pattern classification is introduced, which integrates the ResNet residual network architecture and the Convolutional Block Attention Module (CBAM). To enhance interpretability, we employ Grad-CAM visualizations to highlight salient regions contributing to the model’s predictions, providing insights into the learned features and guiding feature selection for traditional classification methods. Experimental results show that our method achieves a high accuracy of 98.5% on a comprehensive plantar pressure dataset. The approach, proves to be reliable and effective for categorizing plantar pressure distribution patterns, holding promise for clinical applications in foot assessment and treatment.