Interpretable Real-Time Surgical Skill Assessment Via Optical Neuroimaging - Project Summary/Abstract Surgical skill is currently measured via structured checklists or composite metrics based on time and errors, which are typically assessed by experienced proctors. The need for experienced proctors for surgical skill assessment creates a time-consuming workflow and potentially subjective scoring of performance. To solve this, there has been interest in finding a method to assess surgical skill performance without the need for proctors. One approach involves the use of neuroimaging to provide performance assessment and there has been a variety of literature on the use of fMRI and EEG for surgical skill assessment. These modalities come with key drawbacks such as restrictions on head and body motion (fMRI) or poor spatial resolution (EEG). For these reasons, functional near-infrared spectroscopy (fNIRS), a modality with relatively low obstruction of body movement and high spatial resolution was selected as the means for surgical skill assessment. Currently, use of fNIRS in real-time applications is inhibited by current processing practices in the field, due to the prevalence of signal components which originate from outside of the brain. Most fNIRS studies rely on post-hoc processing to remove confounding artifacts and noise. To provide unsupervised surgical skill assessment, an online method of removing noise would be required, however, current aims to provide online noise removal, particularly noise related to superficial physiology, are reliant on short, well-characterized tasks. While data-driven approaches such as deep learning show promise, training large models requires large amounts of labeled data, which are challenging to acquire. To address this Aim 1 will focus on developing a method of generating synthetic fNIRS data using Monte Carlo simulations of photon propagation informed by amortized simulation-based inferencing to generate data. This synthetic data will be used in Aim 2.1 to train a deep learning algorithm for near real-time superficial physiological contribution removal from fNIRS data, allowing us to create an online denoising pipeline. This approach will rely on multiple open-access fNIRS datasets along with surgical skill assessment datasets to ensure the developed method is task-agnostic, allowing for its use in a wide variety of tasks. Finally, in Aim 2.2 we will provide a deep learning algorithm for predicting a widely used metric of surgical skill assessment, the FLS score (developed for the widely accredited Fundamentals of Laparoscopic Surgery program), using fNIRS recordings of surgical task completion, allowing for objective, online surgical skill assessment. Over the course of this project, we will publicly release the data generator and denoising pipeline to be freely used by the field. With Dr. Intes’ expertise in optical imaging, he will provide regular feedback and input on research progress with almost daily interactions. Similarly, weekly interactions with Dr. De regarding deep learning, Dr. Radev regarding Bayesian inferencing and Dr. Cavuoto on surgical skill assessment and neuroergonomics will assist with the development of Condell’s research skills for the proposed project and his future career as a researcher.
