<p>In minimally invasive surgery, the detection of surgical instruments is the core of many operating room assistive systems. However, traditional methods mostly rely on fully supervised learning, which requires the use of large-scale annotated datasets, and high-quality annotation in the medical field is extremely costly. In recent years, weakly supervised learning has shown great potential in medical image analysis, as it only requires image-level annotation to analyze surgical images. This paper introduces weakly supervised object detection into the field of surgical instrument detection, training the model using only image-level annotation to achieve accurate identification of surgical instruments, effectively alleviating the problem of scarce annotation data. We propose a cross-image scenario-aware module that utilizes the similarity of background tissues between different images to suppress background features and replaces the traditional multi-class label loss with contrastive loss to enhance the accuracy of the attention map; we design a novel feature library construction method and a unique instance optimization architecture, effectively reducing information loss during the optimization process. Experimental results show that our model achieves 60.4% mAP and 64.6% mCorLoc on the m2cai16-tool-locations dataset, demonstrating the effectiveness and practicality of the proposed approach for minimally invasive surgical instrument detection.</p>