SynthStrip: Skull-Stripping for Any Brain Image (2203.09974v2)

Abstract: 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.

• The paper introduces SynthStrip, a learning-based skull-stripping method using synthetic data that works robustly across any brain imaging modality.
• SynthStrip outperforms other methods (like BET and ROBEX) across modalities beyond structural MRI, achieving high accuracy (e.g., >96% Dice) with precise boundaries.
• SynthStrip is available open-source within FreeSurfer, making it a practical tool for generalized brain extraction in diverse neuroimaging analyses.

SynthStrip: A Comprehensive Approach to Skull-Stripping for Diverse Brain Imaging Modalities

Skull-stripping is a crucial preprocessing step in neuroimaging analysis, facilitating accurate analysis by excluding non-brain tissues from MRI data. Classical methods, tailored for specific image types and acquisition properties, struggle to accommodate the wide variety of clinical imaging protocols. Classical methods, while well-established, such as those within tools like FreeSurfer, BET, and ROBEX, rely heavily on predefined assumptions about spatial features and are often optimized specifically for T1-weighted images. Consequently, these methods falter when applied to images acquired outside of these parameters, such as those with different MRI contrasts or thick-slice protocols prevalent in clinical settings.

SynthStrip, introduced in this paper, represents a robust, learning-based approach to skull-stripping that addresses these limitations by leveraging synthetic data for training. Contrary to traditional techniques that require high-fidelity training datasets representative of target image types, SynthStrip utilizes a synthetic dataset generated from anatomical segmentation, enabling it to generalize effectively across diverse real-world imaging scenarios. By synthesizing an array of MRI scans that include a broad spectrum of anatomical, contrast, and artifact variations, SynthStrip minimizes reliance on actual acquired image data during training, thus enhancing its adaptability across different acquisition schemes and modalities.

The experimentation revealed noteworthy findings. SynthStrip significantly outperforms traditional and other deep learning-based skull-stripping methods like BET, ROBEX, and DMBE across a wide range of contrasts and modalities. Most notably, its robustness extends beyond structural MRI to modalities including MRA, DWI, FDG-PET, and even CT.

SynthStrip's superior efficacy is quantified through several metrics—Dice overlap, surface distances, and volume differences—where it yields statistically significant improvements. For instance, SynthStrip achieved a mean Dice score consistently exceeding 96% across various datasets while maintaining low mean and maximum surface distance errors. This demonstrates its aptitude not only in extracting brain tissue with precision but also in maintaining spatial plausibility and smoothness in brain masks, an area where other methods falter.

Additionally, SynthStrip's model demonstrates consistent brain extraction across time-series data, vital for processing longitudinal and diffusion-weighted imaging sequences. It achieved notably lower discordant voxel percentages than its counterparts, underscoring its reliability in providing stable results across multiple imaging frames.

A differentiating factor in SynthStrip's development is the use of a signed distance transform (SDT)-based loss, which yields smoother brain mask boundaries than conventional Dice loss functions. This choice highlights the importance of focusing on boundary precision and continuity in skull-stripping tasks.

The implications of SynthStrip extend both practically and theoretically in the field of neuroimaging. Practically, SynthStrip simplifies brain extraction across varied imaging protocols, enhancing the accuracy and reliability of downstream neuroimage analyses, from anatomical studies to functional and diffusion MRI analyses. Theoretically, it sets a precedent for the application of synthetic data in developing generalized, robust learning models that can transcend the limitations of real-world data scarcity and variability.

Future directions may explore extending such synthetic data training paradigms to other image processing challenges, such as fetal imaging, and evaluation in wider clinical and research settings. The availability of SynthStrip as an open-source tool within the FreeSurfer package further promotes its integration and usage across diverse neuroimaging applications, fostering continued development and validation in the research community.