An Extremely Fast and Precise Convolutional Neural Network for Recognition and Localization of Cataract Surgical Tools

摘要

Recognition and localization of surgical tools is a crucial requirement to provide safe tool-tissue interaction in various computer-assisted interventions (CAI). Unfortunately, most state-of-the-art approaches are committed to improving detection precision regardless of the real-time performance, which leads to poor prediction for these methods in intraoperative detection task. In this paper, we propose an extremely fast and precise network (EF-PNet) for tool detection that performs well both in intraoperative tracking and postoperative skill evaluation. The proposed approach takes a single sweep of the single network to achieve rapid tool detection during intraoperative tasks, and also integrates densely connected constraint to guarantee a comparable precision for skill assessment. We demonstrate the superiority of our method on a newly built dataset: cataract surgical tool location (CaSToL). Experimental results with a mean inference time of 3.7 ms per test frame detection (i.e. 270 fps) and a mean average precision (mAP) of 93%, demonstrate the effectiveness of the proposed architecture, and also indicate that our study is far superior to recent region-based methods for tool detection in terms of detection speed, surely with a comparable precision.