A Preliminary Study on Autonomous Robotic Colonoscopy via Deep Neural Network

Colonoscopy is the gold standard for diagnosing colorectal cancer and diseases. However, colonoscopy insertion techniques are difficult to master and the rigorous insertion process can be painful for the patient. To address this issue, we develop an autonomously manipulated colonoscopy system based on detected lumen by a navigation architecture. The Resnet34 is utilized as an image feature extractor in the navigation system. The Resnet34-based neural network is trained to predict the target steering point and the collision probability decision making the tip's steering direction and whether to insert or withdraw, respectively. The tip steering direction is not directed to the deepest point, but rather to the direction where the appropriate technique can be implemented for a high safety colonoscopy procedure. In detail, the datasets of 16,322 are prepared with the goal of implementing techniques such as slalom and slide-by used by colonoscopists during colonoscopy. The performance of the prediction results is quantitatively evaluated through RMSE and F1-score, and the results were 0.208 and 0.868, respectively. In addition, we verified that the navigation network operates with high accuracy throughout the entire procedure by evaluating the accuracy of the navigation network during the entire colonoscopy procedure.