Automated Presurgical Language Mapping via Deep Learning for Multimodal Brain Connectivity

Johns Hopkins University, Baltimore MD

Investigators

Linked publications, trials & patents

Abstract

Project Summary/Abstract Approximately 100,000 people in the United States are diagnosed with a primary brain tumor each year. Neu- rosurgery remains the first and most common therapeutic option for these patients with outcomes linked to the extent of tumor resection. However, larger resections also increase the risk for postoperative deficits, particularly in the motor and language areas of the eloquent cortex. Task fMRI (t-fMRI) has emerged as a powerful nonin- vasive tool for preoperative mapping, but these acquisitions are lengthy and cognitively demanding for patients. Moreover, t-fMRI is unreliable if the patient cannot perform the tasks while in the scanner. Our long-term goal is to develop an automated platform for reliable eloquent cortex mapping across a broad patient cohort that comple- ments the existing clinical workflow. The overall objective of this proposal is to design and validate new machine learning algorithms that leverage the complementary strengths of resting-state fMRI (rs-fMRI) and diffusion MRI (d-MRI), which are both passive modalities and easy to acquire. Our central hypothesis is that the combined structural-functional connectivity information in these modalities will enable us to localize language functionality in patients with brain tumors. Our innovative strategy uses recent advancements in deep learning to capture com- plex interactions in the rs-fMRI and d-MRI data that collectively define the language areas. We will evaluate our hypothesis via two specific aims. In Aim 1 we will develop a graph neural network (GNN) that employs specialized convolutional filters to capture topological properties of the connectivity data across multiple scales. Our GNN will be trained in a supervised fashion and evaluated against t-fMRI activations and intraoperative electrocortical stimulation. In Aim 2 we will conduct an exploratory analysis to retrospectively link our GNN predictions to post- operative changes in language functionality. Namely, we hypothesize that patients for whom the surgical path intersects our GNN predictions will experience greater deficits across fine-grained language subdomains. We will also assess the prognostic value of our GNN predictions, as compared to other clinical factors. We anticipate the proposed research will have a transformative impact on surgical planning by helping neurosurgeons to plan more targeted and safer surgeries, thus improving patient outcomes and overall quality of care.