University of Twente Student Theses
Prediction of laparoscopic cholecystectomy procedure duration using artificial intelligence
Kar, N. van de (2022) Prediction of laparoscopic cholecystectomy procedure duration using artificial intelligence.
PDF
2MB |
Abstract: | A cholecystectomy is the procedure of the surgical removal of a diseased gallbladder. Each year, more than 25,000 cholecystectomies are performed by surgeons in the Netherlands. The high volume of the procedure makes it suited for artificial intelligence applications. The aim of this study is the development of an artificial intelligence network that predicts the remaining procedure time for the laparoscopic cholecystectomy based on video data, and updates the estimated remaining procedure time during the procedure based on the progress. The study consists of two parts. The first part is the development a deep learning network that can accurately and objectively classifying the surgical phases of intraoperative laparoscopic cholecystectomy (LC) videos. All 80 LC videos of the publicly available Cholec80 dataset are used as data source, for comparability with other studies. A residual neural network is used as a base-line deep learning network to classify the surgical phases. The classification results are post-processed by a moving window to filter the network output. After classification, the duration of the individual phases is extracted by detecting the phase transitions. In addition, the importance of adequate labelling of surgical video data is investigated. The network performance metrics of the original annotations of the Cholec80 dataset are compared with revised phase annotations, that are defined based on clinical relevance and technical capabilities. The second part consists of the prediction of the remaining procedure time after each surgical phase. The predictions are based on the phase duration, derived from the detected phase transitions by the phase detector. The model performance of linear regression, random-forest regression and support vector regression are evaluated for predicting the remaining procedure time. The residual neural network has a 79.0% accuracy, 80.5% precision, 78.1% recall and 79.3% F1-score for the original annotations and 85.0% accuracy, 86.3% precision, 84.3% recall and 85.3% F1-score for the revised annotations on the test set. The revised annotation performance metrics showed an improvement of 6.0%, 5.8%, 6.2% and 6.0%, for accuracy, precision, recall and F1-score respectively compared to the original. Post-processing of the phase output removed the noisy character but was susceptible to artifacts. TCNs are advised for future research. The regression model accurately predicted the remaining procedure time based on the phase durations of the LC procedure. The random-forest regression model showed to be the best model to predict the remaining procedure time, with an overall RMSE of 8.5 min and R2 of 0.6 on the test set and with a significant difference to almost all linear and support vector regression results. Although these results improve upon the performance stated in previous research, the model did not yield results that are within the defined standards for use in clinical practice. However, further improvements on the network, dataset and learning process, as described in the recommendations, might enable the possibility for clinical implementation. |
Item Type: | Essay (Master) |
Faculty: | TNW: Science and Technology |
Subject: | 44 medicine |
Programme: | Technical Medicine MSc (60033) |
Link to this item: | https://purl.utwente.nl/essays/89561 |
Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Repository Staff Only: item control page