Advanced boundary-enhanced instance segmentation and spatial-temporal transformer models for automated schizophrenic investigation
No Thumbnail Available
Date
2025
Authors
Imarhiagbe, Osasumwen Raphael
University of Lethbridge. Faculty of Arts and Science
Journal Title
Journal ISSN
Volume Title
Publisher
Lethbridge, Alta. : University of Lethbridge, Dept. of Mathematics and Computer Science
Abstract
Accurate segmentation and detection in neuroimaging is essential for advancing clinical understanding and the diagnosis of schizophrenia. This thesis introduces Boundary-Refined Attention Network (BoRefAttnNet), a novel boundary-refined 3D U-Net variant specifically designed for precise segmentation of subcortical brain structures from structural magnetic resonance imaging (sMRI). BoRefAttnNet incorporates multi-scale boundary attention modules that explicitly highlight anatomically critical edges while suppressing background noise, significantly improving segmentation accuracy for small or complex anatomical structures. Evaluations using FastSurfer-processed sMRI data from the publicly available Centre for Biomedical Research Excellence (COBRE) dataset demonstrate that BoRefAttnNet significantly outperforms conventional 3D U-Net baselines in accurately delineating key subcortical structures, including the hippocampus, amygdala, and basal ganglia.
Building upon this enhanced segmentation capability, we further experiment with a Dynamic Spatial-Temporal Transformer Model (DySTTM) to detect schizophrenia by integrating structural and functional MRI (fMRI) modalities. The DySTTM leverages spatial attention to capture anatomical interdependencies from segmented sMRI data and temporal attention to model dynamic brain connectivity patterns from resting-state fMRI. Experimental results indicate that the integration of these multimodal imaging features using DySTTM provides superior diagnostic accuracy and interpretability compared to established models such as 3D ResNet and XGBoost classifiers.
Description
Keywords
boundary-aware segmentation , attention mechanisms , deep learning , multimodal fusion , dynamic spatial-temporal modeling , medical image segmentation , computational psychiatry