Source-linked AI summary
Automated brain extraction of multi-sequence MRI using artificial neural networks
Fabian Isensee, Marianne Schell, Irada Tursunova, Gianluca Brugnara, David Bonekamp, Ulf Neuberger, Antje Wick, Heinz-Peter Schlemmer, Sabine Heiland, Wolfgang Wick, Martin Bendszus, Klaus Hermann Maier-Hein, Philipp Kickingereder
TL;DR
Existing brain-extraction algorithms are often optimized for healthy brains, motivating methods that can handle pathology and heterogeneous MRI data. The paper introduces HD-BET, an artificial-neural-network algorithm, and reports strong performance across MRI sequences and datasets, while the network design has acknowledged limitations in early decision-making and skip connections.
Problem
Brain extraction methods optimized for healthy brains can be challenged by pathological brains and heterogeneous MRI datasets, motivating a more robust approach.
Method
HD-BET uses an artificial neural network to reconstruct brain masks from MRI data, with local structures forwarded from encoder to decoder stages.
Results
Median DICE coefficients exceeded 0.95 for all evaluated MRI sequences, and performance on cT1-w, FLAIR, and T2-w closely replicated that on T1-w.
Takeaways & Limitations
The reported results support HD-BET as a broadly evaluated brain-extraction method across multiple MRI sequences and challenging datasets.
Takeaways & Limitations
The architecture may encourage overly simple early decision-making and forwards local encoder structures to the decoder instead of using lower-layer information through skip connections.
Abstract
from arXiv · showhide
Brain extraction is a critical preprocessing step in the analysis of MRI neuroimaging studies and influences the accuracy of downstream analyses. The majority of brain extraction algorithms are, however, optimized for processing healthy brains and thus frequently fail in the presence of pathologically altered brain or when applied to heterogeneous MRI datasets. Here we introduce a new, rigorously validated algorithm (termed HD-BET) relying on artificial neural networks that aims to overcome these limitations. We demonstrate that HD-BET outperforms six popular, publicly available brain extraction algorithms in several large-scale neuroimaging datasets, including one from a prospective multicentric trial in neuro-oncology, yielding state-of-the-art performance with median improvements of +1.16 to +2.11 points for the DICE coefficient and -0.66 to -2.51 mm for the Hausdorff distance. Importantly, the HD-BET algorithm shows robust performance in the presence of pathology or treatment-induced tissue alterations, is applicable to a broad range of MRI sequence types and is not influenced by variations in MRI hardware and acquisition parameters encountered in both research and clinical practice. For broader accessibility our HD-BET prediction algorithm is made freely available (http://www.neuroAI-HD.org) and may become an essential component for robust, automated, high-throughput processing of MRI neuroimaging data.
Competing interests
The supplied material includes author disclosure statements alongside descriptions of the study’s evaluation metrics, statistical comparisons, tables, and figures.
- Several authors report no activities related to the present article, while others disclose grants, consulting, or lecture fees from named organizations.
- Tables and figures present dataset characteristics, performance distributions, improvement values, representative cases, and comparisons across MRI sequences.
- Brain-extraction performance is evaluated using the DICE coefficient and the 95th-percentile Hausdorff distance, with statistical comparisons reported for HD-BET and competing algorithms.
1. State-of-the-art brain extraction algorithms
The paper reviews conventional brain-extraction methods and compares their design strategies with HD-BET’s intended out-of-the-box use. Existing approaches rely on intensity and morphology, registration and generative models, atlases, or deformable surfaces, with differing support for pathology and MRI modalities.
- BET and related methods segment intensity images, detect uncertain boundary voxels, and remove them through morphological filtering.
- ROBEX combines affine registration, a discriminative brain-boundary model, and a generative plausibility model.
- BEaST uses atlas-based nonlocal segmentation and selects suitable patches from a prior library using summed squared differences.
- MONSTR uses nonlocal patches from atlases containing multiple MRI sequences and was designed to be robust to pathology.
- For comparison, the study applies competing algorithms with standard parameters and no dataset-specific adaptations, while processing EORTC modalities independently.
- The experimental setting is intended to provide a fair comparison because HD-BET is designed for out-of-the-box use and was not optimized for the test datasets.
2. Artificial Neural Network (ANN)
HD-BET uses a U-Net-style artificial neural network that combines semantic encoding with spatial reconstruction to produce brain masks from multi-sequence MRI.
- Preprocessing: The input data are resampled to isotropic 1.5x1.5x1.5 mm³ spacing and normalized through z-scoring before prediction.The predicted brain mask is linearly upsampled to the original resolution for evaluation.
- Architecture: HD-BET uses an encoder–decoder network with skip connections to combine semantic and spatial information for brain localization.The encoder aggregates semantic information, while the decoder reconstructs spatial information; skip connections transfer feature maps between them.
- Architecture: The architecture uses pre-activation residual blocks in the encoder, with instance normalization and leaky ReLU nonlinearities.Residual blocks learn nonlinear residuals added to their inputs and support deeper architectures and gradient flow.
- Architecture: 128x128x128-voxel patches at 1.5x1.5x1.5 mm³ provide broad contextual information for reconstructing masks when large brain regions are missing.The authors specifically describe traumatic brain injury and resection cavities as such conditions.
- Optimization: Auxiliary loss layers train downsampled segmentations with separate loss terms to address overly local early decisions and slow learning in lower U-Net layers.The global loss is computed as a weighted sum of the auxiliary loss terms.
B) Supplementary Tables
The supplementary material reports performance across MRI sequences and datasets using DICE coefficient and Hausdorff distance, with median DICE values above 0.95 for all sequences.
- Statistical analyses: The supplementary analyses report descriptive statistics, normality tests, and nonparametric tests comparing brain-extraction methods across datasets.Friedman tests were used for NFBS and LPBA40, while the Skilling-Mack test was used for EORTC-26101 and CC-359.
- Cross-validation: Supplementary Figure 1 evaluates HD-BET on pre- and postcontrast T1-weighted, FLAIR, and T2-weighted sequences in the EORTC-26101 training set using five-fold cross-validation.The reported metrics are DICE coefficient and 95th-percentile Hausdorff distance.
- Performance summaries: Median DICE coefficients exceeded 0.95 for all evaluated MRI sequences.DICE values are interpreted as better when higher.
- Sequence comparisons: Performance on cT1-w, FLAIR, and T2-w closely replicated performance on T1-w for both DICE coefficient and Hausdorff distance.Higher DICE and lower Hausdorff distance indicate better performance.