ABC-YOLO: Automated skin burn depth classification using YOLO architectures

TL;DR

This paper introduces ABC-YOLO, a deep learning system that uses YOLO architectures to automatically classify skin burn depths with high accuracy.

Contribution

The study introduces the use of YOLOv11x-seg for skin burn classification, achieving state-of-the-art performance on a multi-source dataset.

Findings

The YOLOv11x-seg model achieved an F1-Score of 0.87 and mAP@0.5 of 0.91 for burn classification.

The model outperformed other YOLO versions and demonstrated strong generalizability across datasets.

Statistical analysis confirmed the significance of the YOLOv11x-seg results.

Abstract

Accurate classification of skin burn depth is vital for determining appropriate treatment and accelerating the healing process. This study conducts a comparative analysis of YOLO-based deep learning architectures for the automated classification of skin burns. Analyses were performed on a robust, multi-source dataset created by combining a proprietary collection of 358 retrospective images from Karadeniz Technical University Farabi Hospital with two large public datasets from Roboflow Universe and Kaggle. All images were meticulously labeled into four burn degrees by expert general surgeons. To enhance model performance and generalizability, various data augmentation and preprocessing techniques were applied. Segmentation-based versions of YOLOv8 and YOLOv11 with different architectural sizes (medium, large, extra-large) were evaluated using metrics such as precision, recall, F1-Score, and mAP. The findings revealed that the YOLOv11x-seg model demonstrated marked superiority over all other tested architectures, achieving an F1-Score of 0.87 and a mAP@0.5 of 0.91. Statistical analysis confirmed the significance of these results. The study demonstrates that the YOLOv11x-seg architecture offers significant potential as a rapid and objective decision support tool in clinical settings. This work makes an original contribution to improving burn diagnosis by integrating a state-of-the-art deep learning model into medical image analysis.

Figures