Deep learning-based recognition of hand anatomical structures and traumatic lesions in surgical images: A proof-of-concept study

BACKGROUND: Computer vision in surgery remains limited by the complexity of anatomical environments and the scarcity of annotated datasets, particularly in hand surgery. Semantic segmentation could enable precise recognition of fine anatomical structures and lesions directly from surgical images. METHODS: A dataset of hand photographs was created from operative and cadaveric dissection images. Anatomical structures and lesions were manually segmented at the pixel level. A DeepLabV3+ model with a ResNet-50 encoder was trained using Dice loss. Segmentation performance was evaluated using Dice scores and confusion matrix analysis. RESULTS: The database contained 710 labeled images (71 000 after data augmentation). The model achieved high segmentation performance for major anatomical structures, including tendons (Dice score: 0.84), muscles (0.80), arteries (0.74), nerves (0.70). Tendon and arterial lesions were also successfully identified. Lower performance was observed for underrepresented structures such as veins, surgical tools, and nails. CONCLUSION: This proof-of-concept study demonstrates the feasibility of deep learning-based semantic segmentation of hand anatomical structures and traumatic lesions. The proposed dataset and model represent an initial step toward computer vision-assisted hand surgery, with potential applications in surgical education, augmented reality, and intraoperative guidance.