Preliminary comparison of FreeSurfer segmentation algorithms in the Wake Forest community‐based cohort and potential impact on ATN classification

TL;DR

This study compares two FreeSurfer algorithms for brain MRI segmentation and finds that newer deep learning methods like SynthSeg provide more reliable results, especially in areas affected by aging and disease.

Contribution

The study evaluates the impact of FreeSurfer segmentation algorithms on brain atrophy classification and highlights the benefits of deep learning-based methods like SynthSeg.

Findings

FreeSurfer v7.2 recon-all produced smaller and more variable volume and thickness estimates compared to SynthSeg.

Cortical thickness estimates in key dementia-related regions showed poor to moderate agreement between pipelines.

SynthSeg improved segmentation robustness in poor-quality scans and reduced the need for manual correction.

Abstract

Acquisition and participant‐related artifacts (atrophy, motion) can degrade the quality of acquired images resulting in poor segmentation of tissue compartments. This can bias estimates of brain volume and thickness used quantify age and disease‐related atrophy, a problem particularly salient in clinical populations. In some cases, poor quality scans may be discarded or repeated incurring additional costs.

Participants (n = 624; cognitively normal [CU;n = 330]; mild cognitive impairment [MCI;n = 214]; dementia [DEM;n = 75]; otherwise not classified [OTHER;n = 5)] enrolled in the Wake Forest ADRC Clinical Cohort (Table 1). Structural T1‐MRI scans were processed using FreeSurfer (v7.2) recon‐all and FreeSurfer (v7.4) recon‐all‐clinical (SynthSeg) pipelines. Tau‐PET (FTP) images were acquired; global, meta‐temporal, and entorhinal tau‐PET (white+gray matter; SUVr) was quantified. Measures of (1) cortical thickness: entorhinal cortex, inferior temporal lobe, temporal pole, and meta‐temporal; (2) volume: gray matter, white matter, and hippocampi; and (3) tau deposition (global and entorhinal SUVr) were compared between pipelines and by cognitive status (Figure 2). GLMs (R2=shared variance) and gaussian‐mixture modeling (cohort‐specific atrophy cutpoints) were performed.

Overall, FreeSurfer (v7.2) recon‐all tended to undersgement, producing smaller volume and thickness estimates (and a wider range of estimates), as compared to FreeSurfer recon‐all‐clinical (e.g., SynthSeg; Figure 1a). For cortical thickness, we observed poor‐to‐moderate associations in signature regions for age‐related dementias including: temporal pole [R2: Left=3%; Right=7%]); inferior temporal lobe (R2: Left=34%; Right=24%), entorhinal cortex (R2: Left=32%; Right=27%), posterior cingulate (R2: Left=40%; Right=39%), and precuneus (R2: Left=39%; Right=36%). Conversely, volumetric estimates were largely comparable across pipelines, except for hippocampi (R2: Left=65%; Right=64%), where we observed a modest drop in agreement for classification of atrophy (N; Figure 1b). ∼13% (hippocampal volume) and 36% (meta‐temporal cortical thickness) of cases were discordant when classifying atrophy (Figure 2: cognitive status). Quantification of tau‐PET (global/regional deposition) was not impacted.

Volumetric estimates were comparable across FreeSurfer segmentation algorithms in our cohort; however, cortical thickness estimates were impacted contributing to discrepant classification of atrophy. SynthSeg segmentations were robust to scan quality (not shown) and recovered susceptible regions (e.g., temporal pole). Deep learning‐based segmentation algorithms (e.g., SynthSeg) require less processing time and may ultimately reduce the need to discard poor quality scans or perform manual segmentation.

Figures