The MLOps team is hyper-focused on automatic clinical workflows through AI, leading the research of cutting-edge AI models in the healthcare field, specifically for projects that impact patient care. The team has extensive experience building and training new models from scratch, with the end-goal of deployment into the clinical space. The MLOps team approaches each project in a unique way, curated to the characteristics of the task at hand to push performance and usability to their absolute limits.
Craniosynostosis Triaging System (CTS)
CTS involves building a triaging system for detecting Infant Craniosynostosis using 3D photography images. Severe cases of Craniosynostosis require surgical intervention, and if caught before 4 months of age can result in a much less intense operation to correct. However, given the sheer number of referrals to the Plastic Surgery Clinic at The Hospital for Sick Children, a majority of these cases miss this window, resulting in a much more invasive operation. The teams’ mission is, therefore, to build a classification system to identify Craniosynostosis within the incoming queue of patients, and to triage the patients thought to be at risk to the front of the line.
The GoNoGoNet project involves building and deploying an edge-computing web server within UHN that enables the real-time segmentation of a laparoscopic surgery feed. A very lean U-net model was trained to predict safe & dangerous areas of dissection within a live surgery, which is to be used to guide surgical trainees and residents away from preventable injuries. The entire end-to-end pipeline was optimized for both speed and performance and is now able to achieve remarkable results delivering inference at 60fps with a latency under 100ms. The next steps include expanding the study for use outside of UHN via cloud computing, in an effort to democratize surgical expertise without the end-users needing custom-built, expensive infrastructure in place.
Pneumothorax Detection Model
The team is working within the Radiology department at UHN to build and deploy a Pneumothorax detection model, which will use x-rays to triage patients for intervention. Currently, due to the often large backlog of x-rays waiting to be read, patients suffering from Pneumothorax can quickly deteriorate while waiting for a radiologist to read and confirm their scan results. Such a triaging system has the potential to catch these at-risk patients and allow the clinical care team to provide the appropriate intervention much quicker.