Feasibility of deep learning for renal artery detection in laparoscopic video / 中华泌尿外科杂志

ABSTRACT

Objective:To explore the feasibility of deep learning technology for renal artery recognition in retroperitoneal laparoscopic renal surgery videos.Methods:From January 2020 to July 2021, the video data of 87 cases of laparoscopic retroperitoneal nephrectomy, including radical nephrectomy, partial nephrectomy, and hemiurorectomy, were retrospectively analyzed. Two urological surgeons screened video clips containing renal arteries. After frame extraction, annotation, review, and proofreading, the labeled targets were divided into training set and test set by the random number table in a ratio of 4∶1. The training set was used to train the neural network model. The test set was used to test the ability of the neural network to identify the renal artery in scenes with different difficulties, which was uniformly transmitted to the YOLOv3 convolutional neural network model for training. According to the opinion of two senior doctors, the test set was divided into high, medium, and low discrimination of renal artery and surrounding tissue. High identification means a clean renal artery and a large exposed area. For middle recognition degree, the renal artery had a certain degree of blood immersion, and the exposed area was medium. Low identification means that the exposed area of the renal artery was small, often located at the edge of the lens, and the blood immersion was severe, which may lead to lens blurring. In the surgical video, the annotator annotated the renal artery truth box frame by frame. After normalization and preprocessing, all images were input into the neural network model for training. The neural network output the renal artery prediction box, and if the overlap ratio (IOU) with the true value box was higher than the set threshold, it was judged that the prediction was correct. The neural network test results of the test set were recorded, and the sensitivity and accuracy were calculated according to IOU.Results:In the training set, 1 149 targets of 13 videos had high recognition degree, 1 891 targets of 17 videos had medium recognition degree, and 349 targets of 18 videos had low recognition degree. In the test set, 267 targets in 9 videos had high recognition degree, 519 targets in 11 videos had medium recognition degree, and 349 targets in 18 videos had low recognition degree. When the IOU threshold was 0.1, the sensitivity and accuracy were 52.78% and 82.50%, respectively. When the IOU threshold was 0.5, the sensitivity and accuracy were 37.80% and 59.10%, respectively. When the IOU threshold was 0.1, the sensitivity and accuracy of high, medium and low recognition groups were 89.14% and 87.82%, 45.86% and 78.03%, 32.95%, and 76.67%, respectively. The frame rate of the YOLOv3 algorithm in real-time surgery video was ≥15 frames/second. The false detection rate and missed detection rate of neural network for renal artery identification in laparoscopic renal surgery video were 47.22% and 17.49%, respectively (IOU=0.1). The leading causes of false detection were similar tissue and reflective light. The main reasons for missed detection were image blurring, blood dipping, dark light, fascia interference, or instrument occlusion, etc.Conclusions:Deep learning-based renal artery recognition technology is feasible. It may assist the surgeon in quickly identifying and protecting the renal artery during the operation and improving the safety of surgery.