A unifying approach to machine learning driven medical image segmentation

点击翻译按钮获取中文摘要

Medical images contain a plethora of information about patient anatomy and disease. They can be roughly broken into two categories based on intended use: diagnostic, or intervention. Medical image segmentation, or contouring, plays a key role in both medical research and clinical care but is very time consuming and costly. A trained expert uses specialized software to manually delineate the boundary of a structure an image, e.g., the heart in a computed tomography (CT) image. A single CT image is a volume of upwards of 256 image slices with perhaps 10 or more structures to segment in each. In research, segmentation is a vital first step to quantify the shape, and appearance of a structure in an image or series of images over time. However, diagnostic imaging is not routinely contoured for clinical care, thus greatly limiting its research potential without expensive contouring. For example, to answer a seemingly simple question, does COVID-19 cause a change in the shape of the heart or lungs in patients pre-and post- infection across a dataset of 1,000 patients could require upwards of 256,000 contours and 700 hours of expert labor (assuming 10 seconds per slice) creating a significant barrier to larger studies with tens of thousands of patients. To unlock the full research potential of Canada's vast repositories of diagnostic medical images automated segmentation technologies are required. Segmentation can also play a key role in clinical care. In diagnostic images automated segmentation could be used to better quantify diseased areas. In contrast, many medical procedures (for example radiation therapy for cancer patients) require segmentation for intervention planning. In these clinical workflows automated segmentation could improve efficiency enabling greater throughput without a cost increase, benefiting Canadians through reduced wait times and reduced healthcare system costs. To date due to technical limitations medical image segmentation methods have been limited to training on purpose-built research datasets consisting of a few hundred patients, with a heavy limit on the number of possible structures, typically 9 or fewer for a given model. We propose to advance state-of-the-art methodologies for medical image segmentation by increasing their generalizability to building models from thousands of routine clinical radiation therapy images, and more than 77 distinct structures. This research will not only advance knowledge in medical image segmentation methodologies but unlock the greater potential of datasets in Canada and develop more robust segmentation models brining us closer to a solved solution for segmentation to use in both research and clinical workflows.