top of page
Machine Learning at the FNNDSC

How is Machine Learning (ML) being used at the FNNDSC?

​

​

The FNNDSC uses various Machine Learning (ML) approaches, in particular Large Language Models (LLMs), Large Visual and Language Models (LVLMs), Conventional Machine Learning approaches, and Deep Learning Neural Networks or Transformers, across a variety of research and development efforts. We also develop AI tools adaptive to application scenarios. 

 

Example focus areas include outcome prediction in patient cohorts (neonatal brain injury, epilepsy, rare neurological diseases, malnutrition), image analysis (segmentation, registration, lesion detection, age prediction, outcome prediction, image and non-image integration), and fetal image reconstruction. Our utilization and development address how to rapidly deploy AI/ML pipelines for a range of possible medical informatics domains.

​​

​

​​

 

Who are contact points for ML at the FNNDSC​

​

 

Dr. Eleonora Tamilia is using predictive AI and ML to advance the diagnosis and treatment of epilepsy. Her work focuses on AI-driven tools that integrate EEG, MRI and clinical predictors to identify epileptogenic brain regions, predict surgical outcomes, and assist in patient diagnosis. Through these efforts, Dr. Tamilia’s group works towards supporting clinical decision-making and improving outcomes for individuals undergoing evaluation and treatment for epilepsy.

​

Dr. Kiho Im is using deep neural networks to improve fetal image processing and analysis. Novel deep learning models have been developed for an automated processing pipeline for fetal brain MRI, including image quality assessment, brain and tissue compartment segmentation, and cortical parcellation. His work also focuses on developing generalizable and explainable ML models, fetal brain age prediction and a generative model of normative fetal brains, to rapidly analyze fetal brain MRI slices as acquired, provide quantitative information on the severity and extent of abnormalities, accelerate clinical interpretations, and improve prognostic models. His work also includes advanced graph neural networks, such as graph prototype and diffusion models, for detecting abnormalities in cortical sulcal patterns.

​

Dr. Rudolph Pienaar leads the development of generic ML training and deployment infrastructure to be used in a wide variety of problem settings, including models of segmentation on brain images, continuous training infrastructure, lung CT/radiograph models, and more. Part of his focus is on integrating ML/AI directly into existing clinical systems.

​

Dr. Yangming Ou’s group (Image, Informatics, and Intelligence, or I3 group) is developing AI/ML algorithms for studying brain development in normal and disease, and to integrate multi-modal clinical data for improved healthcare experience. Applications include: atlas construction; longitudinal atlas-based development quantification; age/sex/IQ/neurocognition outcome prediction; abnormality detection and lesion segmentation in medical images; medical image registration and segmentation; multivariate feature selection when combining MRI, demographics, socioeconomic status, genetic, lifestyle, behavior, environment, and nutritional data for multi-omics learning; disease-specific chatbot; automatic parcellation, understanding, summarization, and standardized digitalization of clinical and radiology reports; AI summarization of large-volume (thousands or more) of documents for knowledgebase construction; foundation AI models to deal with small sample size problem;  LLM/LVLM-based agentic AI for medical diagnosis and prognosis; and medical image, clinical data or report harmonization (dictionary generation, multi-site standardization, converting from unstructured to structured data). The tools that have been developed in this group have been applied to normal lifespan brain development (0-97 years of age), and also in diseases suchs hypoxic ischemic brain injury, sturge-weber syndrome, congenital heart diseases (brain-heart association), 22q11 deletion syndrome, birth cohort nutrition-brain association, and malnutrition studies. Many of these AI tools are based on international cohorts, or consortia the center is building, which encompass >30,000 normal controls 0-100 years of age, and disease populations from dozens of international sites.

​

​​

 

bottom of page