Automated Brain Tissue Segmentation on CT guided by MRI: Advancing AI‐based Neuroimaging for Dementia

TL;DR

This paper introduces a new AI method to segment brain tissues in CT scans using MRI data, improving dementia diagnosis in resource-limited settings.

Contribution

The study introduces MedNeXt, a novel 3D segmentation model that outperforms nnU-Net in multi-orientation and multi-modal brain tissue segmentation on CT.

Findings

MedNeXt achieved higher Dice Similarity Coefficients (DSCs) and volumetric similarity scores compared to nnU-Net across different orientations.

MedNeXt showed better generalizability in dementia cohorts with higher DSC and volumetric similarity scores.

nnU-Net is more resource-efficient, suitable for limited-resource settings, while MedNeXt excels in accuracy and scalability.

Abstract

Brain tissue segmentation is vital in Alzheimer's and dementia research for creating detailed neuroanatomical maps, diagnosing early‐stage neurodegeneration, and guiding interventions. Although MRI remains the standard approach for its superior soft‐tissue contrast, CT is a more accessible imaging modality in acute and resource‐constrained settings.

This study utilized paired CT‐MRI datasets from the Gothenburg H70 Birth Cohort (N = 733) and the Memory Clinic Cohort of the National University Hospital, Singapore (NUS Dementia Cohort, N = 210) to train and evaluate advanced segmentation models—nnUNet (2D & 3D models for 300‐1000 epochs) and MedNeXt (3D‐ Small, Base, Medium and Large models for 3x3x3 & 5x5x5 kernels). MRI‐derived labels were employed to guide CT segmentation, allowing accurate delineation of brain tissue segmentation (Gray Matter: GM, White Matter: WM and Cerebrospinal Fluid: CSF). Evaluation was conducted on all axial datasets for all variations of the models and for coronal & sagittal orientations the best performing models were utilized for inference.

The 3D nnU‐Net achieved average Dice Similarity Coefficients (DSCs) of 0.82, 0.72, and 0.76 for axial, coronal, and sagittal orientations, respectively, while MedNeXt demonstrated slightly superior performance with DSCs of 0.83, 0.73, and 0.78. MedNeXt also exhibited improved volumetric similarity in axial datasets, with scores ranging from 0.842 (CSF, sagittal) to 0.992 (WM, axial). When applied to dementia cohorts, MedNeXt achieved higher generalizability with an average DSC and volumetric similarity of 0.73 and 0.912, compared to 0.70 and 0.854 for nnU‐Net. Extended training (1000 epochs) enhanced nnU‐Net's performance, yet MedNeXt displayed superior scalability, handling larger kernel sizes and multi‐modal imaging scenarios. However, significantly longer training times of up to 288 hours was required for the largest model.

Automated CT brain segmentation guided by MRI‐derived labels demonstrates clinically acceptable segmentation performance on untrained dementia cohort. nnU‐Net is more resource‐efficient and suitable for limited‐resource settings, while MedNeXt has higher accuracy excelling in multi‐orientation and multi‐modal datasets. These findings validate the feasibility of using CT imaging with advanced segmentation frameworks to develop accessible neuroimaging tools for Alzheimer's and dementia research, addressing diagnostic challenges across diverse clinical contexts.

Figures