Gamma Knife Radiosurgery (GKRS) is a technique where patients are treated with high doses of radiation to very small targets in the brain with extreme accuracy and precision, while minimizing exposure to surrounding healthy tissue. MRI is essential for GKRS because it provides the detailed anatomical and functional information needed to define the target and critical brain structures to avoid. Ideally, a patient MRIs would be sufficiently clear, but in practice, the quality of the images can vary substantially. One common problem is patient motion, voluntary or involuntarily, which can cause artifacts in the resulting images.
Currently, evaluation of whether an image is of sufficient quality is mainly a subjective determination by the treatment team, generally after an imaging session has closed. A blurry MRI means sending patients for a repeat scan, adding cost and delaying critical treatment. A real-time tool which could make an objective determination if an image is too degraded would potentially save time and cost in clinical practice.
Researchers Alexandra Ferentinos, Samantha Johnson, and Peter Landers worked on a capstone project where they researched models and techniques to automatically assess if MRI scans have sufficient clarity to be useful for GKRS. The models considered included support vector machines and convolutional neural networks.
Their work showed that machine learning can help with this task; we expect to refine the models more before employing them in the clinical arena. Expanding the training dataset and allocating more computational resources should improve progress towards desired performance benchmarks and make brain surgery even safer.
Researchers: Alexandra Ferentinos, Samantha Johnson, Peter Landers
Sponsors: David Penberthy, David Schlesinger, Jason Sheehan
Advisors: Abbas Kazemipour, Adam Tashman