Flatfoot disorder recognition based on the YOLO-ChA algorithm

We address automatic flatfoot recognition from plantar-pressure images by proposing YOLO-ChA, a task-aware integration of channel attention into YOLO. Unlike generic YOLO + attention variants, YOLO-ChA injects channel attention only in the PANet neck at the P3 and P4 scales (stride 8 and 16) before feature concatenation, in order to preserve arch-relevant low–mid spatial cues. We further introduce an anatomy-guided channel grouping (heel/midfoot-arch/forefoot) that biases feature fusion toward clinically meaningful arch loading patterns without requiring a separate spatial attention branch. This is implemented using a lightweight ECA-style cross-channel interaction that keeps parameters and FLOPs essentially unchanged. On a clinically acquired plantar-pressure dataset of 978 images from 489 subjects (flatfoot vs. normal), YOLO-ChA achieved an overall classification accuracy of 95.14% under subject-disjoint train/test splits and 10-fold cross-validation. Comprehensive ablation studies over attention placement (backbone/neck/head) and attention type (SE/CBAM/ECA-style) show that the proposed neck-only, stride-aware design provides the best trade-off between accuracy and computational cost. These results indicate that domain-informed attention placement can improve flatfoot screening performance in clinical plantar-pressure analysis while remaining lightweight enough for practical deployment.