Publication

Abstract : Background and Objective: Prolonged hospitalization and frequent follow-up visits add financial and mental health instability in wound patients along with an increase in the doctor’s workload. This study uses 2D RGB wound images captured by smartphones to classify wounds using lightweight Deep learning (DL) models into three stages based on the healing capacity. Methods: A novel Indian wound database, Amrita_wound, along with open datasets AZH and Medetec, is precisely annotated into the three grades. Ten lightweight models deployable in mobile or edge devices, applying the concept of transfer learning, are used in the classification of wounds. The models are also perturbed with enhanced features to boost their prediction capabilities. The statistical significance of perturbation is explored. A cross-corpus analysis is incorporated to evaluate the generalizability and robustness of the models. Grad-CAM, an Explainable AI (XAI) tool, supports all classifiers to increase the transparency of the predictions. Results: A performance of above 80% in precision, F1-score, specificity, and sensitivity is reported across all the grades for both MobileViT X S and FastViT T8 models in the multi-class classification of the Amrita_wound dataset. However, the more compact MobileViT X S outperformed with an average sensitivity of 90% and 89% for other measures. In addition to this, upon testing the unseen data Medetec in a cross-corpus (combination of AZH and Amrita_wound) model, an average F1 score of 68% across all grades is achieved, with 77% for Grade 3. Conclusion: The automated wound grading system can aid patients in interpreting the extent of healing and support doctors in prioritizing wound care based on visible severity features. Grad-CAM images enhance the trust of doctors and patients in depending on AI models by evaluating the integrity of model predictions.