Grant: Discovery Grants Program - Individual

Context: In medical imaging, state-of-the-art AI algorithms rely on supervised learning, a paradigm that requires access to labeled data annotated with extensive human labor and expertise. Therefore, the performances are limited by the scarcity of clinical experts available to annotate images other than for patient care. Learning from limited amounts of annotated data makes the models susceptible to failure whenever the data characteristics change. Hence, generalization across the spectrum of pathologies and image-acquisition techniques remains a major challenge for wider adoption. This calls for bringing current algorithms to a whole new level of automation, scalability and accuracy, leveraging the available large-scale amounts of unlabeled images and the uncertain (noisy) knowledge that might be associated with them (e.g., text). Objectives: This research program focuses on novel mathematical models and computational methods for semantic segmentation and classification, the two key problems in medical image interpretation. Following on from our expertise, we intend to pursue novel optimization-based formulations, which leverage large-scale amounts of unlabeled images with medical knowledge. Such a knowledge could be leveraged from various sources, without labeling burden, e.g., from the diagnostic text reports and unstructured clinical annotations stored everyday with the images; from anatomical constraints; or from the relevant clinical literature. Specific technical objectives are: (i) defining novel constraints, which embody domain knowledge, thereby mitigating the lack of data annotations; (ii) designing approximation-based strategies for optimizing the ensuing problems; and (iii) evaluating and validating our investigations on medical images, for their high variability and complexity, the challenges they bring to learning algorithms, and their promises for advancing healthcare practices and research. While this research can serve a breadth of clinical applications, we intend to prioritize domains of high impact on the economy and society, such as oncology. Significance: With formulations integrating domain knowledge and large-scale unlabeled data, as well as advanced optimization and algorithmic expertise, this research program promises to deliver internationally competitive algorithms (in terms of automation, precision, speed and robustness), and to attract high-caliber HQPs and provide them with cutting-edge training in a high-demand sector. The scope goes beyond medical imaging applications. It has potential to impact strongly emerging subjects in AI, with the current paradigm shift towards unsupervised large-scale foundation models. In medical imaging, the potential is strong given the rich domain knowledge. In this application domain, powerful models promise to impact health care research (e.g., the understanding of complex diseases) and practices (e.g., disease early detection, diagnosis, monitoring, treatment and follow-up).