A Robust YOLOv8-Based Framework for Real-Time Melanoma Detection and Segmentation with Multi-Dataset Training

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computational Biology

Background:

• Melanoma diagnosis relies on subjective interpretation, leading to errors.
• Accurate and timely diagnosis is critical for improving melanoma patient survival rates.
• Existing diagnostic methods lack consistency and efficiency.

Purpose of the Study:

• To develop a robust deep learning framework for real-time melanoma detection and segmentation.
• To enhance the generalizability of melanoma detection models across diverse clinical conditions.
• To improve diagnostic accuracy and computational efficiency in melanoma identification.

Main Methods:

• Utilized a YOLOv8-based deep learning framework for unified detection and segmentation.
• Employed a multi-dataset training strategy with ISIC 2020, HAM10000, and PH2 datasets.
• Applied adaptive contrast enhancement, artifact removal, CutMix, and Mosaic augmentation.

Main Results:

• Achieved state-of-the-art performance with 98.6% mAP@0.5, 0.92 Dice Coefficient, and 0.88 IoU.
• Outperformed conventional models like U-Net, DeepLabV3+, Mask R-CNN, SwinUNet, and SAM.
• Demonstrated real-time inference speeds of 12.5 ms per image.

Conclusions:

• The YOLOv8 framework offers high accuracy and efficiency for melanoma diagnosis.
• Multi-dataset training is crucial for robust model generalization.
• Integrating explainable AI can enhance clinical trust and interpretability.