Automatic Detection of Laparoscopic Videos Distortion Using Machine Learning Classification
Abstract
Purpose: Ensuring excellent video quality is crucial for the success of minimally invasive surgical procedures without disrupting the surgical procedure flow. Real-time laparoscopic video frequently encounters issues such as blur and smoke, often stemming from lens contamination. The automatic detection of these distortions is imperative to assist surgeons, ultimately reducing operative time and mitigating risks for the patient.
Materials and Methods: In this paper, we leverage the Laparoscopic Video Quality (LVQ) database developed by Khan et al. to train and validate our model. To classify defocus blur, motion blur, and smoke in the laparoscopic video, we adopt a novel approach utilizing a cascade support vector machine (SVM) classifier, which combines decisions from three binary classifiers. The first classifier categorizes videos into two classes: good and distorted. The second classifier focuses on detecting smoke and blur, while the third is dedicated to distinguishing between defocus blur and motion blur.
Results: In this study, we calculate performance metrics, including accuracy rate, precision, recall, F1 score, and execution time, which are crucial indicators for evaluating quality detection results. The machine-learning classification demonstrates notable performance, with an accuracy rate of 96.55% for the first classifier, 100% for the second, and 99.67% for the third classifier. Additionally, the classification achieves a high inference speed of 37 frames per second (fps).
Conclusion: The experimental results showcased in this paper underscore the efficacy of the proposed approach in automatically detecting distortions in a laparoscopic video. The method exhibits high performance, excelling in both accuracy and processing speed. Notably, the method's advantage lies in its simplicity and the fact that it does not necessitate high-performance computer hardware.