Deep-learning-based tractography for surgical planning in epilepsy treatment

• Grant number: 17/23747-5
• Support type: Scholarships abroad - Research Internship - Master's degree
• Effective date (Start): January 10, 2018
• Effective date (End): July 09, 2018
• Field of knowledge: Engineering - Biomedical Engineering - Medical Engineering
• Principal researcher: Roberto de Alencar Lotufo
• Grantee: Oeslle Alexandre Soares de Lucena
• Supervisor abroad: Sebastien Ourselin
• Home Institution: Faculdade de Engenharia Elétrica e de Computação (FEEC). Universidade Estadual de Campinas (UNICAMP). Campinas , SP, Brazil
• Research place: University College London (UCL), England
• Associated to the scholarship: 16/18332-8 - Deep learning for brain structures segmentation in MR imaging, BP.MS

Abstract

Epilepsy is a neurological disorder that affects over 50 million people worldwide. Among these patients, between20 and 40% of focal epilepsy patients are refractory to antiepileptic medications and are potentially curable with brainresection. Brain epilepsy surgeries usually involves many risks, such as visual field deficit caused by damaging theMeyer's loop of the optic radiation in anterior temporal lobe resections. Therefore, reducing the risk of neurologicaldeficit is fundamental to surgical planning. Tractography is a visualization of the white matter fibers or tracts; its goalin presurgical planing is simply to identify the position of eloquent pathways, such as the motor, sensory, and languagetracts to reduce the risk of damaging these critical structures. Although many tractography methods have been presentedin the literature, there are still outstanding challenges. These challenges include crossing fibers configurations, prematurestopping of fibers, and the trade-off among ad hoc constraints of signal modeling, for instance thresholding according tofractional anisotropy, which local diffusion model and what parameter are settings to use. The most robust methods arestill using many manual steps to ensure anatomically plausible fiber bundles. Deep learning (DL) methods take advantageof the neighborhood information to disentangle asymmetric fiber patterns and there is no trade-off in ad hoc constraintsof signal modeling to overcome due to the ability of DL methods to learn important features from raw data. Furthermore,DL prediction is usually fast which is suitable for on-the-fly approaches for surgical planing. The goal of this researchinternship project is to develop and validate deep-learning-based tractography for surgical planing in epilepsy treatment.