Learning laparoscopic video shot classification for gynecological surgery

摘要: Videos of endoscopic surgery are used for education medical experts, analysis in research, and documentation everyday clinical life. Hand-crafted image descriptors lack the capabilities a semantic classification surgical actions video shots anatomical structures. In this work, we investigate how well single-frame convolutional neural networks (CNN) shot gynecologic work. Together with manually annotate hours raw videos showing endometriosis treatment myoma resection over 100 patients. The cleaned ground truth dataset comprises 9 h annotated material (from 111 different recordings). We use well-known CNN architectures AlexNet GoogLeNet train these both, anatomy, from scratch. Furthermore, extract high-level features weights pre-trained model Caffe zoo feed them to an SVM classifier. Our evaluation shows that reach average recall .697 .515 structures respectively using off-the-shelf features. Using GoogLeNet, achieve mean .782 .617 respectively. With achieved is .615 .469 action main conclusion our work advances general methods transfer domain gynecology. This relevant as natural images, e.g. it distinguished by smoke, reflections, or limited amount colors.