Analysis of surgical artifacts of laparoscopic images as part of the 3D modeling of an automated optic cleaning device

Abstract

During minimally invasive surgery, the laparoscope lens is often contaminated by surgical artifacts such as blood, smoke, and fog. This project aims to develop a software solution for automatically detecting these artifacts to trigger the cleaning of the laparoscope’s optical surface and propose a hardware solution for optimal cleaning. Concerning the software, two artificial intelligence-based models are developed: a machine-learning method using a cascaded support vector machine and a deep learning model with ResNet-50. The Laparoscopic Video Quality database is used for training, testing, and validating the results. The deep-learning approach demonstrates superior accuracy, achieving detection rates of 97.5% for defocus blur, 97.5% for motion blur, and 85% for smoke. However, the machine-learning approach excels in inference speed, reaching 37 frame by second (FPS), making it better suited for real-time applications on low-cost systems. For the hardware component, we used Computer-Aided Design (CAD) to develop a 3D model of the laparoscopic lens-cleaning device; featuring four nozzles arranged in two pairs positioned 120 degrees apart around the lens. Each pair includes two side-by-side nozzles, one for a physiological saline-based cleaning liquid and the other for a CO₂-based drying gas. A 2D cleaning simulation confirms the efficiency of the proposed prototype.