Comparative examination of YOLO models for classification and utilization in nipple injuries
Related Experiment Videos
Ruyi Zhang1, Yang Shu2, Sijie Ding1
1School of Nursing, Jilin University, Changchun, Jilin, China.
Introduction:
Nipple injury is a significant health issue that affects breastfeeding rates globally. Early diagnosis is essential for successfully preventing nipple injury and enhancing breastfeeding rates. The current approach of detecting nipple injuries mostly relies on the subjective assessments of healthcare practitioners. This approach has many disadvantages, including significant subjectivity, limited efficiency, and insufficient diagnostic consistency. This project aims to create an automated diagnosis technique that is objective and successful. YOLO served as the basis for developing an automated approach for detecting and classifying nipple injuries.
Methods:
The preliminary dataset included 613 images of nipple injuries collected from nursing moms in the obstetrics department of the First Hospital of Jilin University between March 2024 and July 2025. The total was increased to 2,452 photos by the use of data augmentation techniques. Utilizing this dataset, we conducted training and assessment on six distinct iterations of the YOLO model: YOLOX , PICOdet, YOLOv5, RTDETR, YOLOv8, and YOLOv11.
Results And Discussion:
The performance studies indicated that YOLOv11 achieved a detection accuracy of 0.85 for nipple injuries, markedly above the performance of other models (PICOdet: 0.78; YOLOv5: 0.75; RTDETR: 0.68; YOLOv8: 0.72; YOLOX: 0.74). Moreover, YOLOv11 exhibited superiority in critical measures, including Precision (P: 0.85) and recall (R: 0.82). To strengthen comparative completeness, representative two stage detectors (Faster R-CNN and Cascade R-CNN) were further included as baselines, and YOLOv11 remained superior under a unified COCO-style IoU evaluation protocol.
