Source-linked AI summary
SynthStrip: Skull-Stripping for Any Brain Image
Andrew Hoopes, Jocelyn S. Mora, Adrian V. Dalca, Bruce Fischl, Malte Hoffmann
TL;DR
Skull-stripping is important for neuroimage analysis, yet classical and supervised learning methods often generalize poorly beyond familiar MRI contrasts and acquisition properties. SynthStrip trains a CNN on highly varied synthetic images generated from anatomical label maps, and the authors report robust extraction across diverse image types with substantial improvements over baselines. The paper also identifies fetal MRI as a future extension area involving thick slices and motion.
Problem
Classical and supervised deep-learning skull-stripping methods are often limited to particular MRI contrasts, resolutions, acquisition schemes, or training-data types.
Method
SynthStrip trains a convolutional neural network using synthetic images generated from anatomical label maps with varied anatomies, contrasts, artifacts, and geometric properties.
Results
SynthStrip substantially outperforms baseline methods across diverse evaluation datasets and achieves the highest Dice score and lowest mean surface distance for more than 80% of test images.
Takeaways & Limitations
A single model can extract brains across imaging modalities, anatomical variability, and acquisition schemes without requiring real target acquisitions during training.
Takeaways & Limitations
The authors identify fetal head extraction from in-utero MRI as future work because fetal scans involve thick slices and between-slice motion.
Abstract
from arXiv · showhide
The removal of non-brain signal from magnetic resonance imaging (MRI) data, known as skull-stripping, is an integral component of many neuroimage analysis streams. Despite their abundance, popular classical skull-stripping methods are usually tailored to images with specific acquisition properties, namely near-isotropic resolution and T1-weighted (T1w) MRI contrast, which are prevalent in research settings. As a result, existing tools tend to adapt poorly to other image types, such as stacks of thick slices acquired with fast spin-echo (FSE) MRI that are common in the clinic. While learning-based approaches for brain extraction have gained traction in recent years, these methods face a similar burden, as they are only effective for image types seen during the training procedure. To achieve robust skull-stripping across a landscape of imaging protocols, we introduce SynthStrip, a rapid, learning-based brain-extraction tool. By leveraging anatomical segmentations to generate an entirely synthetic training dataset with anatomies, intensity distributions, and artifacts that far exceed the realistic range of medical images, SynthStrip learns to successfully generalize to a variety of real acquired brain images, removing the need for training data with target contrasts. We demonstrate the efficacy of SynthStrip for a diverse set of image acquisitions and resolutions across subject populations, ranging from newborn to adult. We show substantial improvements in accuracy over popular skull-stripping baselines -- all with a single trained model. Our method and labeled evaluation data are available at https://w3id.org/synthstrip.
1 Introduction
Skull-stripping removes non-brain MRI signal and supports reliable neuroimage analysis, but existing classical and supervised learning methods often fail to generalize beyond familiar acquisition types. SynthStrip addresses this limitation by training on deliberately diverse synthetic images to extract brains across modalities, resolutions, and populations.
- Motivation: Skull-stripping removes non-brain MRI signal and supports downstream analyses including registration and major neuroimaging pipelines.Brain-extracted inputs help algorithms focus on relevant anatomy and avoid distracting or non-rigidly deforming tissues.
- Limitations of existing methods: Popular classical methods are tailored mainly to near-isotropic, adult, T1w MRI and may perform poorly on infant or thick-slice clinical scans.Stacks of 2D fast spin-echo acquisitions are given as an example of a challenging image type.
- SynthStrip: SynthStrip is presented as learning to skull-strip brain images across a wide range of acquisitions and modalities with one trained model.Figure 1 illustrates examples of extractions from varied image types.
- Limitations of existing methods: Supervised deep-learning extractors achieve high accuracy on image types represented during training but often deteriorate on unseen characteristics.Building sufficiently diverse acquired training datasets requires substantial human effort, limiting the range of supported data types.
- SynthStrip: SynthStrip adapts synthetic-image training from anatomical label maps, exposing a CNN to deliberately broad anatomies, contrasts, and artifacts without real target acquisitions.The resulting model is designed to remain agnostic to acquisition specifics.
2 Related Work
Prior skull-stripping work includes classical deformable-surface and hybrid methods alongside supervised deep-learning approaches. SynthStrip’s related-work contribution is its synthetic-only training strategy, intended to avoid repeatedly expanding datasets and retraining for new image types.
- Classical methods: Classical skull-stripping methods include deformable mesh approaches that reconstruct a smooth brain-surface boundary.BET is identified as a widely used example distributed with FSL.
- Classical methods: FreeSurfer’s hybrid method combines watershed initialization, surface-mesh expansion, and an intensity-distribution atlas to refine brain masks.The atlas helps prevent outliers during mesh fitting, while erroneous mask voxels are removed during post-processing.
- Deep-learning methods: Supervised deep-learning extraction methods use acquired images paired with ground-truth brain masks derived through classical methods or manual segmentation.These approaches have demonstrated accuracy and speed that often outperform classical counterparts.
- SynthStrip’s contribution: SynthStrip focuses on synthetic-only training rather than architectural optimality, aiming to avoid expanding training data and re-optimizing weights for every new image type.The strategy builds on prior deep-learning brain-extraction work while targeting broader image-type support.
3 Method
SynthStrip trains a 3D U-Net entirely on synthetic head images generated from anatomical segmentations, varying anatomy, contrast, artifacts, and acquisition properties beyond realistic ranges. The network predicts brain boundaries through a signed distance transform, optimized against segmentation-derived targets.
- 3 Method: SynthStrip samples whole-head segmentations and synthesizes gray-scale training images with randomized acquisition characteristics for supervised learning.The target brain mask is derived by merging brain labels from each sampled segmentation.
- 3.1 Synthesis: Synthetic training images deliberately exceed realistic ranges of anatomy, intensity distributions, and acquisition characteristics to promote generalization.The synthesis parameters are sampled from predetermined ranges selected to create broad variation.
- 3.1 Synthesis: Spatial transformations combine random affine changes and nonlinear deformations, after which transformed segmentations define both the image source and brain target mask.The transformations expose the network to varied head positions and anatomical irregularities.
- 3.1 Synthesis: Intensity synthesis assigns random tissue-specific distributions, then adds bias fields, gamma augmentation, cropping, down-sampling, and partial-volume effects.Cropping and down-sampling are each applied during synthesis with a 50% probability rate.
- 3.2 Loss: The network predicts a signed distance transform whose positive interior and negative exterior values are thresholded into a binary brain mask.Training compares the predicted transform with an exact target transform derived from the target brain boundary, with distant voxels down-weighted.
4 Experiments
Experiments evaluate SynthStrip against classical and learning-based baselines across diverse real brain-image acquisitions using overlap, surface-distance, volume, sensitivity, and specificity metrics. SynthStrip generally achieves the strongest accuracy and robustness, including on acquisition types for which baselines are tuned poorly.
- Accuracy across datasets: SynthStrip significantly outperforms baseline methods on nearly every dataset and metric, with no baseline significantly outperforming it on any dataset.The comparison covers Dice overlap, mean and maximum surface distances, volume difference, sensitivity, and specificity.
- Accuracy across datasets: SynthStrip achieves the highest Dice score and lowest mean surface distance for more than 80% of test images, compared with less than 10% for BET.This pattern is reported across the full test set and remains strong for T1w, near-isotropic, adult-brain images.
- Robustness across image types: SynthStrip retains superior performance on non-T1w, thick-slice, and infant images, where many baselines generalize poorly or fail entirely.Figure 6 specifically evaluates baselines that generalize beyond common structural T1w MRI scans; FSW fails entirely for multiple non-T1w subsets.
- Mask properties: SynthStrip consistently combines high sensitivity and specificity, whereas other methods tend to over-segment or under-segment brain masks.ROBEX and BEaST achieve high specificity at the cost of low sensitivity, while FSW shows the inverse tendency.
- Qualitative errors: SynthStrip predictions usually contain only minor over-segmentation, with errors concentrated near the dorsal cortex and eye sockets.For images where it does not achieve the best score, SynthStrip lags the best baseline by −0.53 ± 0.54 Dice percentage points (N = 111) and (0.20 ± 0.18) mm mean surface distance (N = 94).
- Loss comparison: The SDT-based loss produces smoother boundaries than traditional soft Dice loss despite comparable global accuracy.Using Ldice increases maximum surface distance by 6.4 ± 3.2 mm and yields 4.5× higher exposed boundary voxels than Lsdt on 132 isotropic FSM images.
5 Discussion
SynthStrip uses diverse synthetic training images to generalize brain extraction across modalities, anatomy, and acquisition schemes. It improves over baselines across nearly all tested cohorts while avoiding substantial mislabeling, and is released with evaluation data.
- 5 Discussion: SynthStrip is trained on diverse synthetic images to generalize across imaging modalities, anatomical variability, and acquisition schemes.The synthetic training distribution exceeds the realistic range of medical images.
- 5.1 Baseline Comparison: SynthStrip significantly improves skull-stripping accuracy for nearly every tested image cohort, matching baseline performance in the few exceptions.The comparison covers diverse cohorts rather than a single acquisition type.
- 5 Discussion: Varying acquisition characteristics during synthesis promotes robustness to protocols including clinical thick-slice acquisitions and large-voxel images.Blurring simulates partial-volume effects, while randomized resolution supports generalization to these acquisitions.
- 5.1 Baseline Comparison: SynthStrip predicts realistic brain-mask shapes, whereas baselines can over- or under-segment tissue when image features are unexpected or unaccounted for.Representative errors include over-labeling and removal of brain matter; SynthStrip errors are uncommon and typically involve small extracerebral regions.
- 5.3 Model and Data Availability: The tool and labeled evaluation data are publicly available to facilitate further development and testing of robust skull-stripping.The associated code is available through the open-source FreeSurfer package.
- 5.4 Future Work: A future extension targets fetal head extraction, where thick slices, motion, artifacts, and wide fields of view make brain localization and extraction difficult.The proposed application concerns in-utero fetal MRI and super-resolution reconstruction workflows.
6 Conclusion
SynthStrip addresses failures of skull-stripping tools outside isotropic T1w scans by synthesizing varied training images and optimizing a contrast- and acquisition-agnostic network. It produces accurate brain masks across widely varying imaging paradigms.
- 6 Conclusion: SynthStrip targets failures of popular skull-stripping utilities on non-T1w contrasts and stack-of-slices acquisitions common in clinical imaging.The paper identifies accurate skull-stripping as important because it directly affects downstream quantitative analyses.
- 6 Conclusion: SynthStrip synthesizes diverse training images from anatomical label maps and optimizes a convolutional neural network agnostic to MRI contrasts and acquisition schemes.The approach is designed for varying contrast and resolution.
- 6 Conclusion: The resulting tool produces highly accurate brain masks across a landscape of imaging paradigms with widely varying contrast and resolution.This conclusion summarizes the method's supported scope rather than a single dataset result.
ROBEX BET 3DSS BEaST FSW DMBE SynthStrip
The comparison reports skull-stripping accuracy across datasets using mean Hausdorff distance and mean percent volume difference. SynthStrip significantly outperforms most baselines, with exceptions identified by nonsignificant p-values.
- ROBEX BET 3DSS BEaST FSW DMBE SynthStrip: Representative error figures compare each method's over-labeling and brain-matter removal across six individual test scans.White arrows indicate over-labeling, while orange arrows indicate removal of brain matter.
- ROBEX BET 3DSS BEaST FSW DMBE SynthStrip: Mean Hausdorff distance measures skull-stripping accuracy across datasets by comparing computed and ground-truth binary brain masks.Table S1 reports mean values with standard deviations and baseline-versus-SynthStrip p-values.
- ROBEX BET 3DSS BEaST FSW DMBE SynthStrip: SynthStrip significantly outperforms most baselines across datasets, except comparisons with orange p-values where p > 0.05.This pattern is reported for both Hausdorff-distance and volume-difference evaluations.
- ROBEX BET 3DSS BEaST FSW DMBE SynthStrip: Mean percent difference in volume measures the discrepancy between computed and ground-truth binary brain-mask volumes across datasets.Table S2 reports mean values with standard deviations and p-values comparing each baseline with SynthStrip.