Grant: Discovery Grants Program - Individual

Context - This research program aims to minimize annotation efforts in medical imaging with innovations in active learning and statistical shape analysis. Annotation effort is indeed fundamental to medical imaging. The collection of medical images effectively grows faster than what human experts can manually process. Unfortunately, methods for automated analysis rely on arduous manual image annotations to train the underlying learning models, while the time needed to label images is inevitably finite. Concurrently, shape analysis has evolved as an essential tool to characterize anatomy using compact representations. Such computational models could lend expertise towards the images and image areas requiring the utmost attention for annotation. Providing such enabling shape-based technology could, therefore, guide annotation efforts in medical imaging and overall reduce the amount of data experts need to review. Research Directions and Methodology - One promising avenue is to leverage learnable shape models in active machine learning. Shape models indeed capture anatomical variabilities, while active learning finds the most informative data to annotate. Bridging this surprising gap in the field will enable active learning in medical imaging owing to shape-guided annotation models. This research program will develop a new family of shape-based algorithms to locate the most relevant images to annotate, assist image segmentation with minimal interaction and streamline corrective steps. Our research objectives thus plan to: (O1) reduce annotation sets by finding optimal annotation candidates in unlabeled images, (O2) alleviate annotation inputs by locating informative image areas for annotation, and (O3) streamline annotation refinements by providing alternative annotations with minimal corrections. This program will first focus on the challenging problem of cardiac image segmentation, routinely lacking annotations. Our shape-guided annotation models fit our long-term vision of reducing annotation efforts in medical imaging. Impact - The contributions constitute new enabling technologies for annotating imaging datasets. The novel shape-based annotation models will have multiple impacts: Socio-economically, our shape-based models will directly reduce annotation time to free hours of burdening clinical tasks and essentially ground new tools to manage expert-time limitations towards the most informative imaging data. Methodologically, our strategies will enable active learning to find the most promising candidate images, pixels, and refinements, thus streamlining annotation efforts in medical imaging. Scientifically, our models also complement the current advances in semi-supervised learning and unsupervised domain adaptation, feeding their algorithms with the most informative imaging data. Our shape models are, therefore, highly relevant to fields where data grows faster than what experts can process, including in computer vision and machine learning.