Improved Skin Lesion Segmentation in Dermoscopic Images Using Object Detection and Semantic Segmentation

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Accurate AI-based skin lesion classification relies on precise segmentation of dermoscopic images.
• Conventional segmentation methods demand extensive pixel-level annotations, which are costly and susceptible to artifacts like hair and skin markings.

Purpose of the Study:

• To introduce a novel hybrid framework, SAM-enhanced YOLO, for efficient and accurate pixel-level segmentation of dermoscopic lesions.
• To evaluate the performance of SAM-enhanced YOLO against traditional methods and standalone SAM.

Main Methods:

• Integrated the Segment Anything Model (SAM) with the You Only Look Once (YOLO) algorithm to create the SAM-enhanced YOLO framework.
• Compared the hybrid framework against GrabCut, Otsu's thresholding, and SAM-only segmentation, initializing SAM-only at the image center.

Main Results:

• SAM-enhanced YOLO achieved a 28% improvement in Intersection over Union (IoU) (0.738) and a 22% improvement in F1-score (0.833) compared to SAM-only (IoU: 0.578, F1-score: 0.683).
• The framework demonstrated consistent performance across various lesion shapes and contrast levels, showing superior robustness and lower variability than other methods.

Conclusions:

• SAM-enhanced YOLO significantly reduces the need for manual pixel-level annotations, offering a scalable and resource-efficient solution for dermoscopic lesion segmentation.
• This framework has the potential to enhance diagnostic workflows, particularly in clinical settings and areas with limited resources.