By Ian Gormely
Oct 11, 2022
The Fight Tumour Team has won the first Cancer Digital Intelligence (CDI) and Vector Institute collaborative Machine Learning (ML) Challenge for Cancer Image Segmentation. The competition allowed teams to explore the computational limitations of medical image object detection and contouring. Team McIntosh Lab placed second.
For the challenge, teams comprised of UHN and Vector researchers created ML models using deidentified 3D radiological medical images for auto-segmentation of regions of interest, such as tumors, for radiation treatment planning and disease monitoring in head and neck cancer patients. Image segmentation is a tedious manual task typically taking several hours per patient performed by radiologists. AI-assisted auto-segmentation has the potential to reduce the time and effort required for radiographers and radiologists to interpret medical imaging results.
In building and training their models, teams used Princess Margaret Cancer Centre’s RADCURE dataset, made up of medical images from patients being treated for head and neck cancer. The models were evaluated based on the accuracy of the contours as well as their complexity and inference time.
The Fight Tumor team, made up of Jun Ma, Rex Ma, and Ronald Xie, won due to the high score they received for the reliability of their model in segmenting the regions of interest. The McIntosh Lab Team, which included Elyar Abbasi Bavil, Siham Belgadi, Tom Purdie, and Kim Sangwook, came in second place, scoring well for the number of trainable model parameters (its complexity) and model run time per patient (inference time), which were the lowest of the competition. As the winning team, the Fight Tumor team will write a manuscript about their results from the challenge, and present their findings at the Toronto Machine Learning Summit in November alongside Vector Faculty Affiliate and Princess Margaret Cancer Centre senior scientist Benjamin Haibe-Kains, whose team originally curated the RADCURE dataset.
A greater degree of responsible data sharing, while managing privacy risks, was one of the recommendations put forward by the expert advisory panel of the Pan-Canadian Health Data Strategy. The Challenge gave life to the idea that doing so is likely to spur health-improving innovation. The collaboration was also a great opportunity for participants and their teams to showcase their findings and their work to others in the field, encouraging broader conversations surrounding AI and the implications it can have to transform patient care and healthcare delivery.
Congratulations to the winners of the 2022 CDI-Vector ML Challenge, and thanks to everyone who participated. This was a successful challenge The challenge targeted a wider variety of researcher, scientist, and student backgrounds within the UHN and Vector community. We look forward to future collaborations and AI-related projects to continue to help transform healthcare innovation.