Synthèse d’observateurs non-linéaires pour l’évaluation des compétences chirurgicales

Abstract

The outcomes of a patient’s surgery are significantly influenced by a surgeon’s technical skill level. The lack of safety during surgeries is prompting an ever-increasing demand for significant training enhancements in the surgical field. In fact, if a surgeon has poor technical skills during an operation, post-surgical complications may occur. Death may result from these complications. The most common cause of post-surgical complications, including re-operation and re-admission, is surgical technical errors. As a result, the need to improve the technical skill acquisition of surgeons through the development of novel and efficient automated methods is increasing from day to day. In the recent years, machine learning and deep learning techniques have been applied to the assessment of surgical skills, with the goal of providing objective, accurate and efficient eval uations. These methods mostly involve the use of convolutional neural networks (CNNs) and recurrent neural networks (RNNs) to analyze video and kinematic data collected during surgical procedures. By analyzing the patterns and movements of the surgical tools, the system can provide a quantitative assessment of the skill level of the surgeon. Several studies have shown promising results in the use of deep learning for surgical skills assessment, with potential appli cations in both surgical education and quality control. In this thesis, we present a deep learning method based on a encoder-decoder architecture that acts as a nonlinear observer to provide an automated soft surgical skills evaluation based on a continuous scale. While this encoder-decoder architecture represents the main contribution, we also present other deep learning architectures based on Deep Feed Forward Neural Networks (DNNs), Recurrent Neural Networks (RNNs) and Long Short Term Memory Networks (LSTM). In all setups, we consider the surgical skill assessment task as a regression task. We trained and tested different models exclusively on surgical kinematic data recorded from surgeons with different expertise levels and compare the obtained regression results.