An Indian database for grading wound healing and cross-corpus classification using perturbation-based Explainable AI...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Prolonged hospitalization and frequent follow-ups create significant burdens for wound patients and healthcare providers.
• Current wound assessment methods can be subjective and time-consuming, impacting patient care and resource allocation.

Purpose of the Study:

• To classify wound healing stages using lightweight deep learning (DL) models and 2D RGB smartphone images.
• To reduce the financial and mental health strain on patients by enabling remote wound assessment.
• To assist healthcare professionals in prioritizing patient care and managing workload more effectively.

Main Methods:

• Utilized a novel Indian wound dataset (Amrita_wound) and open datasets (AZH, Medetec) for multi-class wound classification.
• Employed ten lightweight DL models with transfer learning, enhanced by feature perturbation for improved prediction.
• Incorporated cross-corpus analysis for generalizability assessment and Grad-CAM for explainable AI (XAI) to ensure prediction transparency.

Main Results:

• Achieved over 80% precision, F1-score, specificity, and sensitivity across all grades for MobileViT X S and FastViT T8 models on the Amrita_wound dataset.
• The MobileViT X S model demonstrated superior performance with an average sensitivity of 90%.
• Cross-corpus analysis on unseen data (Medetec) yielded an average F1 score of 68%, with 77% for Grade 3 wounds.

Conclusions:

• An automated wound grading system using AI can empower patients to understand healing progress and help clinicians prioritize care.
• Explainable AI (Grad-CAM) integration builds trust between patients, doctors, and AI models by verifying prediction integrity.