Research
My primary research interest is in Surgical Data Science. I use machine learning for time series data to improve outcomes in surgery and surgical training, based on sensor information collected in the operating room or simulation environment. Due to the diversity in patients and the cost of adverse outcomes, I advocate for the incorporation of domain knowledge into machine learning for surgery. My work facilitates real-time decision support in the operating room, performance assessment and coaching, and increased surgical efficiency.
I am open to new collaborations with academic, clinical, and industry colleagues.
Projects
Data Augmentation for Time Series
Modern methods for deep learning on time series require huge datasets. In reality, however, time series data can be difficult to collect and organize. Thus, most available time series datasets are are small.
Data augmentation, the process of artificially increasing the amount of available data, has significant added value for deep learning tasks. Data augmentation approaches for time series, however, are lacking. We investigate techniques for data augmentation on time series data. This includes techniques such as: kinematic modelling, oversampling, time series interpolation/extrapolation, evolutionary methods, and sequence generative adversarial networks.
Integrating Deep Learning and Domain Knowledge
Machine learning for healthcare is difficult due to the heterogeneity in patients, disease, and surgical technique. We propose integrating clinical domain knowledge into deep learning methods to promote generalization of our models across patients and across surgeons.
Clinical domain knowledge can be integrated as hand-crafted features, through data curation/augmentation, or within the model architecture. We investigate different approaches for achieving this integration with modern deep neural networks.
Holistic Skills Assessment in Diagnostic Ultrasound
Ultrasound is a clinically useful imaging method because it is inexpensive, safe, and real-time. It is difficult for trainees to master, however, because the images are noisy, the sonographer must simulteneously move the ultrasound probe while interpreting the image, and the image must be mentally translated onto the patient. This necessitates training, which can be supplemented with automated methods.
We seek to provide automatically assessment of ultrasound on several facets, including image acquisition, image quality, and image completeness. This uses imaging data and motion data collected during ultrasound scans. To this end, we deploy spatial, temporal, and spatiotemporal convolutional neural networks.
Workflow Analysis in Ophthalmological Surgeries
Cataract surgery (and other ophthalmological surgeries) are some of the most common surgeries in Canada. Understanding and assessing the surgical process is imperative to enabling real-time decision support and quality assurance in the operating room.
This research focuses on understanding the process by which operators perform surgery on the eye through the analysis of microscope videos. We use 2D and 3D convolutional neural networks for both surgical workflow analysis and quality assurance.
Skills Assessment and Training in Interventions
Medical education in Canada follows a competency-based model, where trainees only progress in the curriculum once they have shown mastery of a skill (rather than based on how much time they have spent working on that skill). This requires regular monitoring of performance and provision of feedback. Expert assessment and feedback can be supplemented with automated assessment and feedback.
This work develops automated methods for performance assessment and coaching during simulation-based training for various image-guided interventions (including central venous catheterization, lumbar puncture, ophthalmological surgery, colonoscopy). We leverage time series classification and sequence modelling methods, combined with domain knowledge from clinical experts.
Vasospasm Detection in CT Images
Cerebral vasospasm after haemorrhage is known to correlate with mortality. CT angiogram can be used to detect vasospasm; however, it is difficult and time consuming to read. Thus, diagnosis is not consistent.
We deploy convolutional neural networks for CT angiogram segmentation to detect vessels that are positive for vasospasm. The deep learning model is informed with domain knowledge and combined with a clinical model based on patient factors to improve performance in detection.
Data-Driven Analysis of Septoplasty Surgery
This project, led by researchers at Johns Hopkins University, focuses on quantitatively understanding several aspects of septoplasty surgery.
List of publications
Software
Perk Tutor
Perk Tutor is an image-guided interventions training platform. It provides sensor data acquisition, anatomical visualization, performance analysis and feedback, and real-time instruction. We have shown its value in training in many applications, with emphasis in ultrasound-guided interventions.
SlicerIGT
SlicerIGT is a software toolkit intended to facilitate development of image-guided interventions applications. The toolkit promotes real-time medical navigation through a robust library of registration, calibration, and visualization methods.
Other projects on my Github account.
