Supervisors info:
Κωνσταντίνος Λουκάς, Αναπληρωτής Καθηγητής, Ιατρική Σχολή, ΕΚΠΑ
Παντελής Καραΐσκος, Καθηγητής, Ιατρική Σχολή, ΕΚΠΑ.
Ιωάννης Σεϊμένης, Καθηγητής, Ιατρική Σχολή, ΕΚΠΑ
Summary:
The benefits of minimally invasive surgery (MIS) are well-established for patients (e.g. shorter recovery time and less chance of infection). However, specialised training is required to ensure that these procedures are performed accurately and safely. Thus, laparoscopic training platforms and virtual reality simulators have been developed. Machine learning algorithms can analyse large datasets to reveal previously unrecognised patterns and expand understanding of technical skill.
The objective of this thesis is to apply machine learning algorithms to classify trainees at the beginning/end of their training in a virtual reality laparoscopic simulator (Lap Mentor). Specifically, the dataset included performance metrics for 6 training sessions (trials) in 3 laparoscopic tasks performed by 23 medical students (138 sessions per task). The three tasks selected are Clipping and Grasping (Task 5), Two-Handed Maneuvers (Task 6), and Cutting (Task 7). For each task, the first/last 3 sessions correspond to the start/end of skill practice (Start/End of Training), respectively. The algorithms applied for skill assessment (Start vs. End of Training) are K-Nearest Neighbors, Logistic Regression, Naive Bayes, Linear Discriminant Analysis, Quadratic Discriminant Analysis, Random Forest and Support Vector Machines. In addition, for each algorithm, the application of two dimensionality reduction techniques were investigated: Principal Component Analysis (PCA) and Fisher's score. The analysis was based on two separate experimental schemes: trial-based (algorithm training was performed at the session level) and subject-based (algorithm training was performed at the trainee level).
The analysis indicates that the the highest accuracy rate was achieved using Support Vector Machine (SVM) with linear kernel, with an accuracy rate of 97.08% for task 5, 97.29% for task 6 and 76.43% for task 7, using PCA. In addition, the study identifies, through Fisher's score, the six best features (performance metrics) for each task and the differences in accuracy rates between the two experimental schemes (difference from 1 to 3%). In conclusion, the use of machine learning algorithms can make a significant contribution to the objective evaluation of surgical skills in virtual reality simulators.
Keywords:
Laparoscopy training, Virtual reality simulation, Machine learning, Classification