Date of Award

Spring 2026

Thesis Type

Rollins Access Only

Degree Name

Honors Bachelor of Arts

Department

Computer Science

Sponsor

Dr. Jasser Jasser

Committee Member

Dr. Serina Al-Haddad

Committee Member

Dr. Valerie Summet

Abstract

Automated segmentation of brain MRI remains challenging due to differences in contrast, resolution, and anatomical detail across imaging datasets. In particular, clinical-resolution in vivo MRI lacks the fine structural detail necessary to accurately capture micro-anatomical boundaries. This study investigates whether improving anatomical label granularity, rather than image resolution, can enhance segmentation performance within a synthetic training framework. We utilize the OpenBHB multi-site MRI dataset in combination with the NextBrain atlas, a histology-informed neuroanatomical atlas providing fine-grained structural labels. A unified 15-class label space is constructed through a remapping process, and the SynthSeg framework is trained using synthetic data generation with domain randomization via the Cornucopia library to simulate variability in contrast, noise, and spatial deformation. Two training configurations are evaluated: a baseline model using coarse anatomical labels and an experimental model using fine-grained labels derived from NextBrain. Results show that the fine-grained configuration improves segmentation performance across anatomically complex structures, particularly in regions such as the hypothalamus. These findings suggest that segmentation performance is strongly influenced by anatomical label fidelity rather than voxel resolution alone. This “segmentation-from-labels” approach provides a scalable pathway for improving anatomical precision in MRI segmentation without requiring higher-resolution imaging or extensive manual annotation.

Download

Available for download on Monday, May 07, 2029

Share

COinS