Design, implementation, and validation of novel multimodal deep learning algorithms for clinical translation of diagnostic and prognostic tools for pediatric brain tumors

Pediatric brain tumors are the second most common cancer in children in Canada and the United States. While death from most of these cancers is rare, affected children commonly experience multiple recurrences requiring different types of therapeutic approaches leading to considerable illness. As standard of care, magnetic resonance imaging (MRI) scans of pediatric brain tumors are regularly collected and examined by radiologists to support a wide range of clinical decision making including diagnosis, prognosis, surgical planning, and assessment of treatment response. MRI is non-invasive and can provide a wealth of information on the tumor.  
Recent advances in multimodal Artificial Intelligence (AI) algorithms combining MRI, radiology reports, and clinical data, have shown potential for (molecular) diagnosis and prognosis of pediatric patients with brain tumors comparable to that of histopathology and genetic markers. These promising results suggest that multimodal AI could be a non-invasive alternative to traditional methods based on histopathological sampling obtained through biopsy or surgery, which are invasive. Despite the promising results, there is a large translational gap between the development of multimodal AI algorithms and their clinical implementation mainly due to lack of reliability and generalizability. 
In this proposed research, we will develop novel methods to address the shortcomings of existing multimodal AI algorithms for pediatric brain tumors to enable clinical translation.