Skin Lesion Classification Using YOLOv11 on the HAM10000 Dataset

Abstract

Skin cancer represents a significant global health concern due to its high mortality rate. Early and accurate detection is crucial but often hindered by the limitations of traditional diagnostic methods. This research applies the YOLOv11 algorithm for skin lesion classification directly from dermoscopic images using the HAM10000 dataset (10,015 images, 7 skin lesion classes). The primary objectives are to evaluate YOLOv11's performance in multi-class classification and assess the impact of data augmentation (rotation, horizontal flipping) in addressing class imbalance. The methodology involved two experiments: training YOLOv11 on the original and augmented datasets and comparing its performance with multi-stage architectures (VGG19 and ResNet50). Five pre-trained YOLOv11 models were tested using accuracy, precision, recall, and F1-score metrics. Results showed the YOLOv11x-cls model trained on the augmented dataset achieved the best performance among YOLOv11 models (accuracy 84.74%, precision 83.94%, recall 84.74%, F1-score 84.06%). However, VGG19 recorded the highest accuracy (89.68%). Data augmentation effectively improved model performance by mitigating class imbalance. This study also indicates that multi-stage architectures perform better in skin lesion classification tasks than single-stage architectures. The key contributions of this research are: (1) a comprehensive performance comparison of YOLOv11 with VGG19 and ResNet50 for skin lesion classification and (2) empirical validation of data augmentation's effectiveness in improving model performance. This study demonstrates that YOLOv11 can achieve competitive performance in skin lesion classification despite not surpassing the performance of multi-stage architectures.