Multimodal Deep Learning for Wound Healing Prediction

Chronic wounds, particularly diabetic foot ulcers, heal unpredictably and impose substantial healthcare costs. Deep learning (DL) systems that integrate complementary data streams could support preclinical studies by assessing wound status and therapy efficacy reliably in small-n studies. We developed a three-branch neural network combining convolutional feature extraction for red-green-blue (RGB) images, infrared thermography, and wound-area measurements to classify healing stage and infer treatment response in diabetic mouse wound model. In descriptive ablation analysis, the model integrating RGB and wound-size inputs achieved higher accuracy than the RGB-only configuration (0.75 vs. 0.70), suggesting that explicit size context provided useful complementary information. Thermal images appeared qualitatively informative during early healing but contributed limited additional signal later, likely reflecting subtle inflammatory gradients or angiogenesis in healthy murine skin; this observation motivated a constrained weighted-fusion strategy. In the development-stage comparison, static fusion with tunable fixed weights yielded the highest observed performance (∼0.85 accuracy on the development test subject). Grad-CAM visualizations showed that predictions arose from biologically relevant regions rather than image artifacts. Experiments without subject-level partitioning produced inflated metrics, underscoring the need for grouping by subject to ensure generalizable evaluation. Nested Leave-One-Group-Out cross-validation with KerasTuner-based hyperparameter tuning yielded a mean accuracy of 0.71, estimating performance on unseen subjects. This resulting proof-of-concept model predicts treatment responses for wounds treated with nanofibrillated-cellulose hydrogel. Multimodal DL combining standard and thermal photography with size metrics can stage wound healing and detect divergent treatment trajectories and thus be used in preclinical efficacy evaluation of novel wound treatments. With larger, clinically annotated datasets, this approach could be applied in a clinical setting as well.