Semi-supervised Surgical Tool Detection Based on Highly Confident Pseudo Labeling and Strong Augmentation Driven Consistency

Surgical tool detection in computer-assisted intervention system aims to provide surgeons with specific supportive information. Existing supervised methods heavily rely on the volume of labeled data. However, manually annotating location of tools in surgical videos is quite time-consuming. To overcome this problem, we propose a semi-supervised pipeline for surgical tool detection, using strategies of highly confident pseudo labeling and strong augmentation driven consistency. To evaluate the proposed pipeline, we introduce a surgical tool detection dataset, Cataract Dataset for Tool Detection (CaDTD). Compared to the supervised baseline, our semi-supervised method improves mean average precision (mAP) by 4.3%. In addition, an ablative study was conducted to validate the effectiveness of the two strategies in our tool detection pipeline, and the results show the mAP improvement of 1.9% and 3.9%, respectively. The proposed dataset, CaDTD, is publicly available at https://github.com/evangel-jiang/CaDTD.