DINOv2-Enhanced Surgical Instrument Segmentation with Lightweight CNN

Accurate segmentation of surgical instruments is critical for the development of intelligent surgical systems, especially in automated and robot-assisted surgeries. This paper presents the DINO-DeepSurg model, which combines DINOv2 with DeepLabV3+ to leverage DINOv2's global semantic modeling capabilities while retaining the efficiency of CNNs in local feature extraction. This integration significantly improves the segmentation accuracy of DeepLabV3+, particularly in handling instrument occlusion and fine-grained features in complex surgical scenarios. Experimental results show that the model achieves an average mIoU of 93.10 on the MICCAI 2017 EndoVis Challenge dataset, outperforming existing methods. The addition of the SEBlock module further enhances the model's ability to process fine-grained features, improving segmentation accuracy at complex backgrounds and instrument boundaries. Overall, the DINO-DeepSurg model provides an efficient and accurate solution for surgical instrument segmentation, demonstrating great potential for clinical applications. Future work will focus on optimizing the model's inference speed and adapting it to more complex surgical scenarios to further advance its clinical deployment.