Exploring Feature Contrast for Weakly Supervised Instrument Segmentation in Robotic Surgery

Accurate segmentation of surgical instruments is crucial for robot-assisted minimally invasive surgery, as it enhances intelligent scene awareness and ensures operational safety and efficacy. Most existing methods address surgical instrument segmentation in a fully supervised manner, relying on extensive pixel-wise annotated datasets that are labor-intensive to produce, limiting both effectiveness and scalability. To address this issue, we introduce a novel weakly supervised approach for surgical instrument segmentation, utilizing image-level labels to generate fine-grained segmentation. We propose TrFC, a Transformer-based model that incorporates feature contrast to enhance feature representation. Specifically, we propose an inter-class feature contrast, which encourages the model to learn class-discriminative features by differentiating the features of different classes. Additionally, an intra-class feature contrast is designed, enforcing the model to generate consistent features under varying image conditions, thereby improving its robustness across diverse scenarios. We extensively evaluate our approach on a robotic surgery dataset, and the experimental results demonstrate that our method achieves superior segmentation results with high accuracy and robustness. This approach offers significant potential to advance the efficiency and scalability of surgical instrument segmentation and holds promise for practical clinical application.