Source-linked AI summary
Robust machine learning segmentation for large-scale analysis of heterogeneous clinical brain MRI datasets
Benjamin Billot, Colin Magdamo, You Cheng, Steven E. Arnold, Sudeshna Das, Juan. E. Iglesias
TL;DR
Clinical brain MRI datasets are much more heterogeneous than the high-quality scans targeted by conventional neuroimaging tools, limiting their automated analysis. SynthSeg+ uses hierarchical, domain-agnostic networks with denoising and quality control to analyse such scans, and demonstrates robustness across clinical acquisitions, including more than 14,000 scans in an ageing study.
Problem
Clinical brain MRI analysis requires robustness across varied contrasts, resolutions, artefacts, and subject populations, whereas existing automated tools are limited to narrower acquisition domains.
Method
SynthSeg+ combines hierarchical domain-agnostic CNN modules, denoising, cortical parcellation, and regression-based quality control for heterogeneous clinical MRI segmentation.
Results
SynthSeg+ improves robustness to clinical artefacts, supports accurate parcellation across tested domains, detects erroneous segmentations, and replicates volume trajectories in more than 14,000 clinical scans.
Takeaways & Limitations
SynthSeg+ supports large-scale quantitative morphometry on heterogeneous clinical data and may enable studies with sample sizes considerably higher than research datasets.
Takeaways & Limitations
Residual smoothing slightly lowers accuracy on high-resolution scans, and future work is needed to quantify performance across pathologies.
Abstract
from arXiv · showhide
Every year, millions of brain MRI scans are acquired in hospitals, which is a figure considerably larger than the size of any research dataset. Therefore, the ability to analyse such scans could transform neuroimaging research. Yet, their potential remains untapped, since no automated algorithm is robust enough to cope with the high variability in clinical acquisitions (MR contrasts, resolutions, orientations, artefacts, subject populations). Here we present SynthSeg+, an AI segmentation suite that enables, for the first time, robust analysis of heterogeneous clinical datasets. In addition to whole-brain segmentation, SynthSeg+ also performs cortical parcellation, intracranial volume estimation, and automated detection of faulty segmentations (mainly caused by scans of very low quality). We demonstrate SynthSeg+ in seven experiments, including an ageing study on 14,000 scans, where it accurately replicates atrophy patterns observed on data of much higher quality. SynthSeg+ is publicly released as a ready-to-use tool to unlock the potential of quantitative morphometry.
Significance Statement
SynthSeg+ addresses the difficulty of analysing heterogeneous clinical brain MRI scans by combining domain-agnostic deep learning with hierarchical processing and automated quality control. It is publicly available for out-of-the-box use and is designed to support scalable neuroimaging analysis.
- SynthSeg+ is presented as the first tool for automated analysis of highly heterogeneous clinical brain MRI scans.It is designed to analyse scans across variable contrasts and resolutions without retraining.
- The tool is publicly released with FreeSurfer and can be run out of the box using a simple command.The documented command enables robust mode through the --robust option.
- Domain-agnostic CNNs are trained to handle MRI scans with different contrasts and resolutions, addressing the domain-gap problem affecting conventional supervised networks.Existing CNNs generally perform well only on training domains with similar acquisition properties.
- Automated quality control is included because robust segmentation methods can still produce inaccurate results, particularly on scans with variable quality.The suite uses predicted quality scores to support rejection of faulty segmentations.
- SynthSeg+ combines hierarchical networks and denoisers to improve robustness to difficult clinical acquisitions and topological errors.Hierarchical models decompose segmentation into easier operations, while denoising methods target implausible or erroneous structures.
Results
SynthSeg+ combines hierarchical CNN modules and denoising with automated quality control to improve segmentation and parcellation across heterogeneous clinical MRI scans. It also supports intracranial volume estimation and large-scale volumetric analyses, while retaining strong performance across difficult acquisitions.
- Whole-brain segmentation: SynthSeg+ uses hierarchical modules, including tissue classification, denoising, segmentation, and cortical parcellation, to decompose the analysis into easier intermediate steps.The denoiser corrects topological mistakes before target segmentation, while subsequent modules recover sharp boundaries.
- Whole-brain segmentation: 76 mean Dice points were achieved across clinical scans, with improvements over SynthSeg of 23.5 points for big fails, 5.1 for mild fails, and 2.4 for passes.SynthSeg+ achieved the best scores in all three visual-QC categories.
- Resolution and contrast robustness: Above 80 Dice points were maintained across tested contrasts and resolutions; SynthSeg+ was superior at lower resolutions by up to 1.6 Dice points.SynthSeg was slightly better at 1 mm isotropic resolution, by up to 1.4 Dice points.
- Cortical parcellation: Cortical parcellation improved by 17.1 Dice points on big fails and 1.8 points on good cases, with SynthSeg+ gaining 2.6 points on 7 mm proton-density scans.SynthSeg remained more accurate at high resolution, with a largest gap of 1.6 Dice points on 1 mm T1-weighted scans.
- Automated quality control: The automated QC approach reached accuracies above 96%, and its regressor obtained scores similar to a dedicated state-of-the-art method without a statistical AUC difference.A simpler comparison of denoiser and segmenter outputs achieved 88.4% accuracy but only 66.8% specificity.
Discussion
SynthSeg+ combines hierarchical segmentation, denoising, and quality control to robustly analyse heterogeneous clinical brain MRI scans. Its experiments support broad generalisation, while identifying residual high-resolution smoothing and areas for future extension.
- Quality control: Highly precise quality control automatically detects and discards the few erroneous segmentations, which mainly arise from very low signal-to-noise ratio or insufficient brain coverage.The regression-based QC strategy identifies the cases where SynthSeg+ fails.
- Robustness: SynthSeg+ improves robustness to clinical artefacts through hierarchical segmentation modules, yielding higher Dice scores and fewer outliers than SynthSeg.The method combines domain-agnostic CNNs with hierarchical processing and denoisers.
- Limitations: Residual denoiser-induced smoothing causes marginally lower accuracy than SynthSeg on high-resolution scans, a setting described as relatively uncommon in clinical practice.The authors consider this limitation minor relative to the robustness gain.
- Cortical parcellation: SynthSeg+ provides volumetric cortical parcellation across heterogeneous clinical acquisitions, maintaining high performance beyond the 1 mm T1-weighted scans supported by existing automated tools.Its parcellation performance remains high across tested domains, including heterogeneous clinical scans.
- Ageing study: More than 14,000 uncurated clinical scans yielded volume trajectories that accurately replicated patterns observed in much higher-quality data.Figure 6 compares all scans, QC-passed scans (N = 13,028), and low-resolution scans with slice thickness greater than 6.5 mm.
- Future work: Future work includes cortical-thickness estimation, multimodal processing, and more precise performance quantification across diverse pathologies.The paper notes that cortical thickness may be challenging at low resolution and that pathology-specific evaluation remains to be developed.
Materials and Methods
SynthSeg+ is evaluated on heterogeneous clinical and multi-contrast MRI datasets using a hierarchical, domain-agnostic segmentation suite with automated quality control. The pipeline is trained on synthetic data and supports segmentation, parcellation, volumetric estimation, and failure detection.
- Training datasets: 1,020 anatomical segmentation maps from OASIS, ADNI, and HCP provide the basis for synthetic training data.The maps contain labels for 31 brain structures.
- Test datasets: 15,346 uncurated MGH clinical scans span diverse ages, contrasts, orientations, and resolutions, while additional test sets include 66 multi-contrast scans and 100 ADNI scans.The ADNI subset contains paired 1 mm isotropic T1-weighted and 5 mm axial FLAIR acquisitions.
- Architecture: Three segmenters and a denoiser decompose whole-brain segmentation into coarse tissue labeling, correction, target labeling, and cortical parcellation.The hierarchy is designed to split segmentation into easier intermediate operations.
- Synthetic training: Domain randomisation generates extremely diverse synthetic uni-modal images by varying generative-model parameters over broad uniform priors.Training examples are created from anatomical label maps using random spatial transforms and image synthesis.
- Automated quality control: A regression network predicts performance scores for 10 representative regions to enable automated detection of faulty segmentations.The module is introduced because heterogeneous clinical scans can produce erroneous predictions.
- Inference and measurements: Soft predictions support regional volume and intracranial-volume estimation, while inference produces hard segmentations after resampling and intensity normalisation.The whole inference pipeline takes 12–16 seconds per scan on an RTX6000 GPU.
Robust machine learning segmentation for large-scale analysis of heterogeneous clinical brain MRI datasets – Supplementary materials
The supplementary materials document the large clinical cohort, examples of segmentation failures, and comparative automated quality-control performance.
- Clinical cohort: 15,346 clinical scans were acquired during 1,367 subject sessions at Massachusetts General Hospital.The cohort’s imaging protocols vary in contrast, acquisition direction, and resolution.
- Segmentation failures: Failed segmentations include missing structures, incorrect anatomical labels, and over-segmentation in scans with low contrast, low signal-to-noise ratio, or very low resolution.These examples were identified by visual quality control.
- Automated quality control: The regression-based quality-control method slightly outperformed Liu et al.’s method, while the denoiser-based strategy performed worst.No statistical difference was found between the regression-based method and Liu et al.’s method.
Supplement 4: Quantitative evaluation of the robustness to different subject populations
SynthSeg+ maintains high segmentation accuracy across age, treating age as a proxy for morphological variability. This stability is observed in 500 clinical scans with ground truth.
- Quantitative evaluation: 500 clinical scans with ground truth were used to evaluate accuracy as a function of age.Age was taken as a proxy for morphological variability.
- Quantitative evaluation: SynthSeg+ maintains a high level of accuracy across the whole age range.
- Quantitative evaluation: Two-sided Wilcoxon signed-rank testing found no statistical difference at p=0.05 across the age-based comparisons reported.