Abstract:  In the field of laparoscopic surgical instrument detection, aiming at the insufficient utilization of multi- scale features in the existing object detection networks and the limitations of the existing detection models in adapting to downstream tasks through fine-tuning methods, a laparoscopic surgical instrument detection algorithm based on the improved Retinanet is proposed. The algorithm includes an adaptive prompt fine-tuning module that promotes the efficient adaptation of the model to downstream tasks, the DR-UNet(Deep Residual UNet) module and the RFA(Residual Feature Augmentation) module that enhance the model’s ability to capture information of targets of different sizes. Experiments on the laparoscopic surgical instrument detection dataset show that the proposed algorithm achieves 1. 1% improvement over the baseline model at an IOU(Intersection over Union) of 0. 5, with an mAP @ 0. 5 of 97. 3% . The detection performance is superior to many other state-of-the-art methods, demonstrating significant value and importance in practical applications such as surgical instrument detection. 

Key words: object detection, laparoscopic surgery, prompt learning, multi-scale features, Retinanet, Transformer, UNet