Enhanced Classification of Laparoscopic Video Distortions with ResNet-18 Architecture

The reduced quality of laparoscopic videos can directly affect a surgeon's visibility and may compromise computational tasks in robot-assisted surgery. Several solutions have been proposed to confront this challenge, focusing on detecting and categorizing distortions found in laparoscopic videos. In this study, we present a Residual Network (ResNet18) for automatically identifying and categorizing noise (,NO‘), smoke (,SM‘), uneven illumination (‘UI’), defocus blur (,DB‘), and motion blur (‘MB’) in laparoscopic videos. Our approach involves initial training on an expanded version of the Laparoscopic Video Quality (LVQ) database (LVQ Challenge dataset). To assess the effectiveness of our methodology, we utilized a total of 20660 frames extracted from the five classes of this dataset. The achieved accuracy scores consistently demonstrated high performance across all classes, with rates of 99.50% in training and 99.15% in validation. These results highlight the model's robust ability to generalize effectively and provide accurate predictions.