ENHANCING BURN SEVERITY DETECTION THROUGH ITERATIVE SEMI-SUPERVISED LEARNING: A PRACTICAL FIRST AID GUIDANCE SYSTEM

Abstract

The visual assessment of burn severity is subjective and critical for determining appropriate treatment. This research presents an end-to-end deep learning system that classifies burn degrees from digital images and provides immediate, actionable first aid recommendations. An initial baseline model, employing a MobileNetV2 architecture with transfer learning, achieved a test accuracy of 76% on a public, cleaned dataset, where performance was primarily limited by the dataset's modest size. To overcome this limitation, a state-of-the-art iterative semi-supervised learning (SSL) strategy using pseudo-labeling was implemented. This approach leveraged a pool of unlabeled data to progressively enhance the model's accuracy over successive training cycles. This methodology resulted in a significant performance increase, culminating in a final peak test accuracy of 87.50%. Detailed analysis of the champion model reveals a robust weighted-average F1-score of 0.88 and an exceptional precision of 96% for identifying critical 3rd- degree burns, highlighting its clinical reliability. This study demonstrates that iterative SSL is a powerful technique for overcoming data scarcity in medical image analysis and validates the creation of a high-performance, practical tool for initial burn assessment.