PaperlayerView exact revision on arXiv

Source-linked AI summary

FastSurfer -- A fast and accurate deep learning based neuroimaging pipeline

Leonie Henschel, Sailesh Conjeti, Santiago Estrada, Kersten Diers, Bruce Fischl, Martin Reuter

arXiv:1910.03866v4eess.IVcs.CVq-bio.NC

TL;DR

Traditional neuroimage pipelines are too slow for large-scale MRI studies because they depend on computationally intensive processing. FastSurfer combines deep-learning whole-brain segmentation with cortical surface reconstruction and thickness analysis, achieving fast processing alongside reported reliability and dementia-group sensitivity.

  • Problem

    Traditional neuroimage pipelines use computationally expensive optimization steps that limit their scalability to large MRI cohorts.

  • Method

    FastSurfer combines a 95-class deep-learning segmentation network with fast cortical surface reconstruction, spherical mapping, and thickness analysis.

  • Results

    FastSurferCNN segments a 3D 1 mm isotropic MRI brain scan in under 1 minute, while the full pipeline performs thickness analysis in below 1 hour.

  • Takeaways & Limitations

    FastSurfer provides a pipeline intended to scale automated neuroimage analysis to large datasets while retaining surface-based analysis and supporting applied dementia-group comparisons.

  • Takeaways & Limitations

    Generalizability beyond image types encountered during training remains uncertain, and quality control is still required when applying FastSurfer to new images.

Abstract

from arXiv · show

Traditional neuroimage analysis pipelines involve computationally intensive, time-consuming optimization steps, and thus, do not scale well to large cohort studies with thousands or tens of thousands of individuals. In this work we propose a fast and accurate deep learning based neuroimaging pipeline for the automated processing of structural human brain MRI scans, replicating FreeSurfer's anatomical segmentation including surface reconstruction and cortical parcellation. To this end, we introduce an advanced deep learning architecture capable of whole brain segmentation into 95 classes. The network architecture incorporates local and global competition via competitive dense blocks and competitive skip pathways, as well as multi-slice information aggregation that specifically tailor network performance towards accurate segmentation of both cortical and sub-cortical structures. Further, we perform fast cortical surface reconstruction and thickness analysis by introducing a spectral spherical embedding and by directly mapping the cortical labels from the image to the surface. This approach provides a full FreeSurfer alternative for volumetric analysis (in under 1 minute) and surface-based thickness analysis (within only around 1h runtime). For sustainability of this approach we perform extensive validation: we assert high segmentation accuracy on several unseen datasets, measure generalizability and demonstrate increased test-retest reliability, and high sensitivity to group differences in dementia.

1. Introduction

Large-scale neuroimaging studies require timely MRI processing, but traditional pipelines rely on costly optimization and lack integrated fast surface analysis. FastSurfer addresses this gap with deep-learning segmentation and a complete cortical surface pipeline.

  • Contribution: FastSurfer integrates volumetric segmentation, cortical surface reconstruction, and morphometric estimation, including cortical thickness.The pipeline is presented as a complete processing alternative rather than a segmentation-only method.
  • Motivation: Traditional neuroimaging pipelines require parameter tuning and expensive numerical optimization, resulting in long runtimes.Examples include nonlinear registration and Bayesian segmentation.
  • Deep-learning architecture: FastSurferCNN segments whole-brain MRI into 95 classes in 1 minute on GPU and approximately 14 minutes on CPU.The architecture is based on three plane-specific F-CNNs with view aggregation.
  • Method: The pipeline combines competitive dense blocks, spatial information aggregation, and fast spherical cortical mapping to retain anatomical context while reducing computational demands.The surface pipeline uses spectral mapping based on Laplace Eigenfunctions and transfers cortical labels from images to surfaces.
  • Validation: Validation covers segmentation accuracy, unseen-dataset generalizability, test-retest reliability, and sensitivity to group differences in dementia.The authors describe this as an extensive validation across publicly available datasets.

2.1. Datasets

The study uses multiple publicly available MRI datasets for training, testing, and validation, with FreeSurfer DKT parcellations providing the segmentation reference framework.

  • Datasets: MRI volumes from eight publicly available datasets were selected for training, testing, and extensive validation.Training used ABIDE II, ADNI, LA5c, and OASIS datasets, with additional datasets used for evaluation.
  • Reference processing: All datasets were processed using FreeSurfer v6.0 and the Desikan–Killiany–Tourville atlas for training and evaluation.The atlas-based labeling was adapted to limit the number of segmentation labels.

2.2. FastSurfer CNN

FastSurferCNN uses three plane-specific 2D networks with competitive architectural components and multi-slice context to segment whole-brain MRI efficiently.

  • Architecture: FastSurferCNN segments 95 brain classes in under 1 minute on GPU and approximately 14 minutes on CPU.Three F-CNNs process coronal, axial, and sagittal slices before final view aggregation.
  • Architecture: Each plane-specific network uses four dense encoder and decoder blocks separated by a bottleneck layer.The blocks include dense connections and skip connections in an encoder-decoder architecture.
  • Competitive dense blocks: Competitive dense blocks replace feature-map concatenation with maxout activations, retaining position-wise maxima and keeping input channels constant.This reduces parameters and memory requirements while inducing competition between feature maps.
  • Competitive skip pathways: Competition is also applied across long-range skip connections before decoder processing.The unpooled decoder information and corresponding encoder features are combined through maxout rather than concatenation.
  • Spatial information aggregation: The network receives seven stacked neighboring slices and predicts the middle slice to combine local 3D context with a global 2D view.The input contains three preceding, current, and three succeeding slices.
  • View aggregation: Separate coronal, axial, and sagittal networks aggregate probability maps to account for the brain’s 3D geometry.The sagittal predictions receive half the weight of the other two views during final aggregation.

2.3. FastSurfer Pipeline

FastSurfer modifies the FreeSurfer surface-processing workflow around a full DKT volume segmentation, omitting obsolete steps and adding rapid spectral surface mapping. The resulting pipeline produces FreeSurfer-like volume and surface measures with substantially reduced runtime.

  • FastSurfer integrates neural-network segmentation with cortical surface reconstruction, thickness maps, and regional summary statistics following the DKT atlas.
  • The pipeline provides a fast alternative to FreeSurfer, with speed-up evaluated alongside reliability and sensitivity of the complete workflow.
  • The pipeline uses full DKT segmentation to create a brainmask, retrospectively performs bias correction and linear Talairach registration, and initializes later processing from these results.
  • Initial cortical surfaces are generated with marching cubes to obtain higher mesh quality using slightly fewer vertices than the traditional approach.
  • FastSurfer maps cortical surfaces to a sphere by embedding Laplace-Beltrami eigenfunctions and projecting the resulting three-dimensional coordinates to unit length.
  • Cortical labels are mapped from the image onto the surface to compute regional thickness and curvature statistics without requiring nonlinear spherical atlas registration.

2.4. Statistics

The analysis evaluates segmentation accuracy, boundary similarity, measurement reliability, and sensitivity to diagnostic-group differences. It uses Dice similarity, average Hausdorff distance, intraclass correlation, and matched statistical models for FreeSurfer and FastSurfer.

  • Segmentation accuracy: Dice similarity coefficient evaluates segmentation performance by comparing binary ground-truth and prediction label maps for each cortical and sub-cortical structure.
  • Segmentation accuracy: Dice similarity is used both to compare network architectures and to estimate agreement between FastSurferCNN and FreeSurfer v6.0 on unseen datasets.
  • Boundary similarity: Average Hausdorff distance evaluates segmentation boundaries, with smaller values indicating better boundary capture and zero representing a perfect match.
  • Reliability: Intraclass correlation coefficient assesses correlation and absolute agreement between measurements, with values closer to 1 indicating higher reliability.
  • Sensitivity to group differences: Group-separability analyses fit identical linear models to FreeSurfer and FastSurfer measurements, explaining thickness or volume by diagnosis while controlling for age and sex.
  • Sensitivity to group differences: Because both methods process the same images and sample sizes, smaller p-values provide a meaningful indication of better group separability.

3. Results

FastSurferCNN achieved high segmentation accuracy across unseen datasets, anatomical structures, scanners, sequences, and disease stages, while supporting surface analysis and clinically relevant dementia comparisons. It also showed strong reliability and substantially faster processing than FreeSurfer, although DSC evaluations depend on the accuracy of automated or manual reference labels.

  • 3.1.1. Comparison to FreeSurfer: 0.6% higher DSC and 5.7% lower AVG HD for subcortical structures, plus 1.9% higher DSC and 12.7% lower AVG HD for cortical structures, were achieved versus QuickNAT.Final DSC values were 89.08 and 85.88, with AVG HD values of 0.1400 and 0.2222, respectively.
  • 3.2. Generalizability: FastSurferCNN maintained high accuracy across eight challenging THP sites and generalized to defaced, downsampled HCP images and multiple scanner vendors.On THP, subcortical DSC was 89.00 and cortical DSC was 86.16; site differences were smaller than for QuickNAT on cortical structures.
  • 3.2. Generalizability: 86.35 right-hemisphere and 86.97 left-hemisphere average surface DSC were obtained across 31 cortical parcels, with no structure below 72.3.The surface results corroborated the volume-based segmentation comparisons.
  • 3.2. Generalizability: 83.17 DSC on subcortical MEF images exceeded FreeSurfer’s 82.37 cross-sequence agreement, while cortical agreement was similar at 75.67 versus 75.70.FreeSurfer’s MPRAGE segmentation served as the reference in this comparison.
  • 3.2. Generalizability: DSC comparisons assume accurate FreeSurfer or human-rater labels, although both reference types can degrade across scanners, sequences, and advanced neurodegeneration.This limits direct interpretation when ground truth cannot be trusted.
  • 3.3. Reliability and sensitivity: FastSurfer showed higher subcortical volume test-retest reliability than FreeSurfer, with average ICC values of 0.99 versus 0.97 across 13 structures.The paper also evaluated sensitivity to dementia-related group differences, including cortical thinning and disease-progression associations.

4. Discussion

FastSurfer combines a deep-learning segmentation network with accelerated surface reconstruction to provide a faster alternative to FreeSurfer, while retaining strong reliability and sensitivity across evaluations. Its applicability is broad, but training-data availability, generalization boundaries, and comparisons with full 3D networks remain important constraints.

  • Contributions: FastSurferCNN combines competitive network architecture and spatial information aggregation for detailed whole-brain segmentation, while the full pipeline accelerates cortical surface processing.Across five datasets, it achieved higher average DSC and lower average AVG HD than FreeSurfer on subcortical and cortical structures.
  • Limitations: The study does not deeply compare 2.5D and full 3D networks, although the optimized 3D UNet did not outperform the view-aggregating 2.5D architectures.The authors identify improved 3D architectures and direct 2.5D-versus-3D comparison as future work.
  • Runtime: Below 1 minute is sufficient for FastSurferCNN segmentation of a 3D 1 mm isotropic MRI scan, enabling rapid applications that do not require surface models.Full cortical-thickness analysis completes below 1 h with parallelization, or 1.6 h including surface registration, compared with 4 h for a parallelized FreeSurfer run.
  • Generalizability: FastSurfer generalized across sites, vendors, field strengths, scanner types, disease groups, and unseen image characteristics, including de-faced, downsampled, and multi-echo scans.The authors recommend visual inspection for each acquisition setting because stable generalization to any T1-weighted sequence is not guaranteed.
  • Validation: FastSurfer exhibits improved test-retest reliability, with average ICC values of 0.92 for cortical and 0.99 for subcortical structures across both hemispheres.It also detected Alzheimer’s-related volume and cortical-thickness differences that were detected with FreeSurfer.
  • Limitations: Training on FreeSurfer-generated outputs enables many training cases but can transfer consistent FreeSurfer segmentation errors to the network.Manually labeled or corrected data are limited because producing them is time intensive.

Appendix

The appendix describes the datasets used for training, validation, testing, and reliability assessment, spanning multiple scanners, field strengths, populations, and acquisition settings.

  • Datasets: ABIDE II contributes 1044 MRI scans from 19 institutions, with participants aged 5–64 years and voxel resolutions from 1.30 mm to 0.7 mm.Twenty ABIDE-II cases were used for training.
  • Datasets: The appendix also lists ADNI, LA5c, MIRIAD, and MMND datasets, covering Alzheimer’s disease, neuropsychiatric disorders, longitudinal aging, and multi-sequence structural MRI.MMND includes structural data from 16 healthy volunteers, while ADNI combines 1.5T and 3T scans from GE, Philips, and Siemens vendors.
  • Datasets: HCP Young Adult contains 1200 healthy participants aged 22–35 years scanned at 0.7 mm isotropic resolution on a customized 3T Siemens Skyra.
  • Reference data: Mindboggle-101 comprises 101 manually corrected labeled brain images with anatomically labeled surfaces and volumes based on the DKTatlas protocol.The OASIS-1 test-retest subset also includes manually labeled subcortical segmentations.
  • Datasets: OASIS-1 and OASIS-2 provide scans from nondemented and demented individuals, including a 20-subject OASIS-1 test-retest component and longitudinal OASIS-2 data.The datasets include acquisitions from 1.5T and 3T Siemens scanners.
  • Datasets: THP contains 3D MP-RAGE scans from 5 healthy subjects acquired at 8 imaging centers to assess multi-site neuroimaging reliability.The sites used Siemens 3T TIM Trio or Philips 3T Achieva scanners.

Accuracy

Table A.2 defines the proposed FastSurferCNN segmentation labels and their correspondence to FreeSurfer labels across subcortical and cortical structures.

  • Label mapping: FastSurferCNN labels are mapped to FreeSurfer labels separately for subcortical and cortical structures.
Loading 1910.03866v4…