Percutaneous Nephrostomy Guidance by a Convolutional Neural Network Based Endoscopic Optical Coherence Tomography System

Percutaneous nephrostomy (PCN) is a commonly used procedure for kidney surgeries. However, difficulties persist in precisely locating the PCN needle tip during its insertion into the kidney. Challenges for PCN needle guidance exist in two aspects: 1) Accurate tissue recognition, and 2) Renal blood vessel detection. In this study, we demonstrated an endoscopic optical coherence tomography (OCT) system for PCN needle guidance. Human kidney samples are utilized in the experiments. Different renal tissues including: 1) cortex, 2) medulla, 3) calyx, 4) fat, and 5) pelvis can be clearly distinguished based on their OCT imaging features. We conduct kidney perfusion experiments to mimic the renal blood flow. Our system can efficiently detect the blood flow in front of PCN needle using Doppler OCT function. To improve surgical guidance efficiency and alleviate the workload of radiologists, we employ convolutional neural network (CNN) methods to automate the procedure. Three CNN models including ResNet50, InceptionV3, and Xception were applied for tissue classification. All of them demonstrate promising prediction results, with InceptionV3 achieving the highest recognition accuracy of 99.6%. For automatic blood vessel detection, nnU-net was applied, and it exhibited intersection over unions (IoU) values of 0.8917 for blood vessel and 0.9916 for background.