Explainable deep learning-based multiclass classification of foot radiographs into normal, plantar fasciitis, and flatfoot.

BACKGROUND

The medial longitudinal arch plays a critical role in foot biomechanics, and its abnormalities are associated with conditions such as flatfoot and plantar fasciitis. Early and accurate diagnosis of these disorders is clinically important, but plain radiographs have limitations in visualizing soft tissue pathology.

PURPOSE: This study aimed to develop and interpret a deep learning model capable of classifying foot radiographs into normal, plantar fasciitis, and flatfoot categories.

METHODS

A DenseNet-121 architecture was trained on 9500 synthetic lateral foot X-ray images from the AI-Hub dataset, with augmentation applied for generalization. Model interpretability was enhanced using Grad-CAM++ to identify class-specific regions of interest, followed by quantitative analysis of six spatial attention features.

RESULTS

The model achieved an overall accuracy of 98.53% on an independent test set, with F1-scores of 0.9900 for normal, 0.9837 for plantar fasciitis, and 0.9823 for flatfoot. Visualization revealed anatomically consistent activation patterns: midfoot and arch regions in normal cases, calcaneal and plantar fascia insertion sites in plantar fasciitis, and inferiorly displaced midfoot-hindfoot regions in flatfoot. Quantitative analysis confirmed significant group differences in activation area, intensity, compactness, spatial focus, and entropy (p < 0.001).

CONCLUSION

These findings demonstrate that explainable deep learning models trained on synthetic radiographic datasets can achieve high classification performance and provide interpretable, anatomically potential insights from structural radiographs. This approach may offer preliminary guidance for developing clinical decision support system, while highlighting the need for further validation in real clinical settings.