Artificial Intelligence-Assisted Visual Mapping of Viable and Necrotic Tissue During Wound Debridement: A Clinical Image Segmentation Study

Background: Accurate distinction between viable and necrotic tissue is essential during wound debridement, but visual assessment may be affected by blood coverage, uneven illumination, blurred tissue transitions, and instrument occlusion. This study developed and evaluated a deep learning-assisted image segmentation method for wound tissue assessment. Methods: A retrospective data set of clinical wound debridement images was annotated at the pixel level as background, viable or preservable tissue, and necrotic or nonviable tissue. Tissue-aware copy-paste augmentation was used to improve necrotic tissue representation. A SegFormer-based semantic segmentation model with boundary-aware output was trained and evaluated using pixel accuracy, intersection over union (IoU), Dice coefficient, precision, recall, and boundary F1 score. Results: The augmented data set contained 356 image-mask pairs, including 156 original and 200 synthesized pairs. Training, validation, and test sets included 249, 53, and 54 samples, respectively. On the test set, pixel accuracy was 98.01%. For viable and necrotic tissue, mean IoU and mean Dice were 85.97% and 92.45%, respectively. Dice scores were 92.24% for viable tissue and 92.67% for necrotic tissue. Necrotic tissue recall was 94.09%, and mean boundary F1 was 61.83%. Conclusions: The proposed method showed strong region-level agreement with expert annotations and high sensitivity for necrotic tissue. Color-coded prediction maps and overlay visualizations may support wound tissue assessment, although exact boundary localization remained less accurate. Larger independent clinical data sets are needed for further validation.