A Robust Approach for Classifying Laparoscopic Video Distortions using ResNet-50

During minimally invasive surgery, the laparoscopic lens can be contaminated by condensation, smoke, blood, and debris, directly impacting both the surgeon's visibility and the quality of the laparoscopic video. To address this issue, various solutions have been proposed for detecting and classifying different types of distortions in laparoscopic videos. Our method focuses on the automated identification and categorization of five distortion categories in laparoscopic videos, as defined in the Laparoscopic Video Quality database by Khan et al. These distortions are noise ('NO'), smoke ('SM'), uneven illumination ('UI'), defocus blur ('DB'), and motion blur ('MB'). Utilizing Residual Networks (ResNet50), our method achieves high prediction rates in detecting and classifying these distortions: 100% for 'SM', 100% for 'NO', 99.82% for 'UI', and 99.92% for 'DB', and 99.87% for 'MB'. These consistently high scores across all categories highlight the model's robust ability to generalize effectively and deliver precise predictions.