Fine-Tuning YOLOv8n for Precision Detection of Blackhead, Fungal Acne, and Acne Nodule

Accurate differentiation among acne subtypes, such as blackhead, fungal acne, and acne nodule, is crucial for effective dermatological treatment. However, it presents a significant challenge for automated visual analysis due to the subtle and overlapping characteristics of lesions. This research aimed to enhance the precision of automated acne detection by meticulously fine-tuning the YOLOv8n object detection model specifically for the three challenging types of acne. A curated dataset comprising 2008 images derived from public sources (Skin-90, DermNet) was utilized. The dataset was manually annotated for blackhead, fungal acne, and nodule by author under the supervision of a domain expert to ensure clinical accuracy. An iterative fine-tuning process systematically explored key hyperparameters, including image size, multi-scale training, and diverse data augmentation strategies, with performance evaluated using Precision, Recall, and mean Average Precision at an IoU of 0.5 (mAP50). The optimal configuration (Experiment 6: imgsz=640, multi-scale training, mixed geometric augmentations with scale=0.5, degrees=10.0, translate=0.15, and cls_loss weight of 0.75) achieved the highest overall precision (P<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">all</inf>) of 0.693 and a robust overall mAP50 of 0.589. Notable per-class mAP50 scores were 0.442 for blackhead, 0.750 for fungal acne, and 0.576 for acne nodule. The qualitative analysis further corroborated the model's capability for precise lesion localization, although challenges in specificity for blackheads (evidenced by some false positives) were noted. This study successfully demonstrates that a systematically fine-tuned YOLOv8n model can achieve high precision in differentiating challenging acne subtypes, offering a promising approach for developing more reliable computeraided diagnostic tools in dermatology,