Neurosurgical Planning with Robust Eloquent area Delineation from Individualized Connectivity-based Techniques (NeuroPREDICT)

Project Summary/AbstractThe idiosyncrasies of the human brain require that individualized mapping of functional regions be performed before surgical interventions for cancer or epilepsy. The success of this mapping procedure has direct effects on surgical outcomes and preserving cognitive and sensory function post-surgery. Current gold standard procedures for pre-surgical mapping are invasive, time-consuming, and technically demanding. Several non-invasive procedures have emerged in recent years; however, they have not yet displaced the gold standard procedures. Task-based functional magnetic resonance imaging (t-fMRI), the most widely used non-invasive pre-surgical mapping technique, requires that patients perform cognitive or motor tasks while in the scanner—a time-consuming and expensive procedure. Also, not all patients can perform fMRI tasks due to language barriers, sensory deficits, being unconscious, etc. Connectome Fingerprinting (CF) is a recently developed technique that uses machine learning to train a model capable of predicting functional brain activation from task-free resting-state fMRI (rs-fMRI). Once trained on a set of t-fMRI and rs- fMRI data, an unseen subjectandapos;s unique pattern of brain activation can be predicted using only an rs-fMRI scan of their brain—therefore eliminating the need to perform tasks during the fMRI scan. Despite the promise of CF, the accuracy of the current best practice modeling techniques is not high enough yet to be clinically useful and studies applying CF have nearly always used healthy populations. Much research remains to be done to increase the accuracy of CF models before they can be deployed for pre-surgical mapping.The long-term objective of the research proposed here is to develop a software application that combines applied machine learning with medical imaging to provide a non-invasive means for mapping the brains of neurosurgical patients before surgery. Importantly, we aim to increase the accuracy of CF modeling by expanding the modeling efforts to probabilistic Bayesian approaches that leverage prior information from the structure of the data. We will test a wide array of tunable data and model parameters to arrive at a current recommendation for best practices in CF research and applications. Finally, we will test our modeling procedures with a dataset of healthy control and pre-surgical patients diagnosed with brain tumors. We will test the softwareandapos;s ability to accurately predict functional brain organization in these patients and adaptively retrain the models to produce the most accurate results. This work has the potential to revolutionize pre-surgical brain mapping and expand its applicability to a greater number of patients.Project NarrativeIn the United States, approximately 24,000 new cases of brain tumors are reported each year, with many patients requiring expensive pre-surgical planning and mapping of functional regions to minimize post-surgical impairments. In many neurosurgical practices [96% per 1], this involves performing a time-consuming and costly task-based fMRI acquisitions (nearly $1200/scan in 2014; [2]) before surgery to identify eloquent brain areas recruited for motor control, language, and cognition that must be spared during surgery. By combing task-based fMRI, resting-state fMRI and advanced machine learning to map the functional topology of the brain, the proposed technology will lower pre-surgical planning costs, reduce the burden on physicians and technicians, and expand pre-surgical mapping to previously excluded patients who cannot perform fMRI tasks.