<p>Scar, as the residual evidence of skin injury, is an important evaluation index in forensic identification. However, due to the subjectivity and errors of manual segmentation and calculation, the identification lacks objectivity and accuracy. Therefore, seeking intelligent and accurate segmentation methods has practical significance in forensic science. This study proposes a scar image segmentation method based on Segment Anything Model 2 (SAM2). Through self-built datasets, the stage-wise fine-tuning strategy is adopted to optimize the mask decoder and prompt encoder. The multi-scale feature fusion and dynamic weight mechanism are introduced to enhance the model’s processing ability for low contrast, small targets, and fuzzy boundaries. Compared with nnU-Netv2, MedSAM-2, and pre-trained SAM2, the fine-tuned SAM2 achieved better performance in indicators such as Dice, IoU, and accuracy, verifying its effectiveness and promotion potential in medical image segmentation tasks.</p>