Transfer learning and Machine Learning Classification for Laparoscopic Video Distortion Detection

Distortions like blur and smoke in real-time laparoscopic videos often result from lens contamination. Detecting these distortions automatically and "in real time" is a step preceding automatic lens cleaning and leads to a clear vision for surgeons, hence reducing surgical time and minimizing risks for patient. Our approach leveraged transfer learning, transposing knowledge from natural images to the laparoscopic video domain. We utilized a pre-trained ResNet50 convolutional neural network (CNN) to extract image features, subsequently processed by a cascade of support vector machine (SVM) classifiers to categorize various distortions. This last strategy amalgamates outputs from two binary classifiers. The first classifier distinguishes videos as good or distorted. The second classifier focuses on smoke and blur detection. The first classifier attains 99 % accuracy. The second classifier achieves 100% accuracy in detecting smoke and blur. These values prove the effectiveness of our approach, which combines ResNet50-based transfer learning and cascaded SVM classification for automatically detecting smoke and blur distortions in laparoscopic videos. Such results are promising for the detection of the remaining distortions.