Healthcare

Summary of Findings

This study proposes an efficient deep learning-based framework for automated flatfoot (pes planus) detection from weight-bearing foot X-ray images. It integrates convolutional feature extraction using a modified VGG16 backbone with rigorous statistical feature selection, reducing the feature space from 128 to 11 discriminative features. Four machine learning classifiers were evaluated.

• The Random Forest (RF) classifier achieved the best overall performance with perfect precision (100.0%) and an AUC-ROC of 1.00 (95% CI: 0.99–1.00) on validation data using the selected 11 features.
• On the unseen test set, RF achieved an accuracy of 98.2%, precision of 100.0%, recall of 95.3%, and F1-score of 97.6%.
• Decision curve analysis (DCA) showed that the RF model consistently achieved the highest net benefit across a wide range of threshold probabilities on both validation and test sets.
• Integration of explainable AI (XAI) using LIME provides local interpretability, highlighting image features contributing to predictions, even in cases with mixed feature contributions or counter-intuitive explanations.

Enterprise Implications

• Automated flatfoot detection can streamline clinical workflows, reducing time-consuming manual assessments and inter-observer variability.
• The use of a compact, statistically selected feature set (11 features) makes the system computationally efficient and easier to interpret, crucial for integration into existing clinical IT infrastructures.
• High accuracy, precision, and net clinical benefit (demonstrated by DCA) indicate strong diagnostic reliability, potentially improving patient outcomes through early and consistent detection.
• Explainable AI (LIME) enhances trust and adoption by clinicians, allowing them to understand the model's decision-making process and validate its reasoning, even when explanations are occasionally counter-intuitive.