Critical analysis of the performance of the YOLO architecture in the detection of oral lesions from clinical images
Gustavo Goetz Ribeiro, Jean Schmith, Rita Fabiane Teixeira Gomes, Giovanna Nunes Machado, Vinícius Coelho Carrard, Rodrigo Marques de Figueiredo · 2025
The objective of this work was to train a convolutional neural network for the detection of oral cavity lesions in heterogeneous clinical images, using YOLOv5, as well as the critical analysis of its effectiveness. The database had four categories of elementary oral lesions, without stipulating protocols for obtaining the images, such as distance, angle, and illumination. YOLO showed the best performance in detecting vesicular/blister lesions in both models analyzed: YOLOv5m mAP@50 was 86.8% and in the YOLOv5x model it was 80.7%, followed by papule/nodule in both tests. Images containing only one lesion showed better performance. We considered the quality of the detections obtained to be satisfactory in the majority of images despite using a small dataset for this evaluation.