Artificial Intelligence-Assisted MRI Screening for Pediatric Cancers

ASPIRE Award (2019-Present)

Anna Goldenberg, PhD (Principal) and Andrea Doria, MD, PhD, SickKids Research Institute; Casey S. Greene, PhD, University of Pennsylvania; and L. J. States, MD, Children's Hospital of Pennsylvania

Whole body magnetic resonance imaging (wbMRI) is an important tool for screening children and adults with genetic, hereditary predispositions for cancer. Artificial intelligence-based tools have already been used successfully to improve detection by wbMRI of early stage cancer in adults. Unfortunately, because it is difficult to obtain and interpret wbMR images of children and also because of the relative rarity of childhood cancers, there are not enough wbMR images of children available to train AI algorithms to detect anomalies. Nevertheless, wbMRI can still be effective for early detection of childhood cancers in at-risk patients, and the right AI tool may be able to improve the rate of early diagnosis. The Goldenberg lab is using a new approach to machine learning called Generative Adversarial Networks (GANs), which involves generating wbMR images that can then be used to train an algorithm. This approach will allow the team to use the relatively limited number of images taken at SickKids over the last several years to generate enough images of children both with and without cancer to create an AI training set. The researchers plan to make these data publicly available so that medical professionals and other researchers can use them. The team will then develop a machine learning model to detect childhood cancers from wbMR images, which will be tested and refined at SickKids to ensure that it can detect cancer in children with a genetic predisposition at a higher rate than current techniques. To test the generalizability of the detection system the team will then use a “federated learning” approach in which data distributed across different centers are used to update models locally. The first instance to deploy this model under the funding of this grant will be with the Children’s Hospital of Philadelphia. The updated models from each institution will then be aggregated into a single tool.  Upon completion of this project, the researchers expect to deploy their tool to other medical institutions and initiate a clinical trial to demonstrate its effectiveness. This tool may one day be in regular use for screening children at risk for cancer, ensuring that they are caught early and allowing for less aggressive treatment and a better prognosis. The utility of this approach may also be expanded to adult cancer patient surveillance, especially for early detection of metastases.