SLANet: 3D Brain Segmentation via Tiling

• SLANet is a deep learning framework that segments the whole brain in 3D by partitioning MRI scans into spatially localized tiles processed with independent 3D U-Nets.
• The method uses affine registration to a canonical atlas and a majority-vote label fusion strategy to ensure consistent, submillimetric, 133-label segmentation.
• SLANet combines classical multi-atlas segmentation techniques with modern deep learning to dramatically reduce computation time while improving segmentation accuracy.

Spatially Localized Atlas Network Tiles (SLANT) is a framework for detailed, high-resolution 3D whole brain segmentation from structural MRI, designed to alleviate the constraints of GPU memory limits and limited training data in the context of deep learning. SLANT divides the volume into overlapping, spatially registered tiles, each processed by an independent 3D convolutional neural network (CNN), typically a modified U-Net, and then fuses their predictions in atlas space via majority voting. This enables submillimetric, multi-label (133-class) segmentation at practical runtimes, leveraging both classical multi-atlas segmentation (MAS) and modern deep learning approaches (Huo et al., 2018, Huo et al., 2019).

1. Architectural Design and Tiling Strategy

SLANT operates by first registering the input T1-weighted MRI to a canonical atlas space (MNI305, $172 \times 220 \times 156$ voxels, 1 mm isotropic). The brain volume in this space is then partitioned into $k$ spatially defined sub-volumes, or "tiles," each indexed as

$$\psi_n = [x_n : x_n + d_x,\; y_n : y_n + d_y,\; z_n : z_n + d_z],\;\; n=1,\dots,k.$$

Two principal configurations have been described:

• SLANT-8: $k=2 \times 2 \times 2=8$ non-overlapping tiles.
• SLANT-27: $k=3 \times 3 \times 3=27$ tiles, each $96 \times 128 \times 88$ voxels, overlapped (≈50% along each axis) so voxels are covered by up to 8 tiles, which improves segmentation consistency at boundaries.

Each tile is processed by an independent, multi-class 3D U-Net. All tiles share the same architecture but learn spatially localized contextual representations due to their fixed atlas-aligned position (Huo et al., 2018, Huo et al., 2019).

2. Registration and Localization

Prior to segmentation, each subject scan $I_{\text{subject}}$ undergoes affine registration to the standard atlas, via a linear mapping:

$x_{\text{atlas}} = A \cdot x_{\text{subject}} + b,$

where $A$ is a $3 \times 3$ matrix and $k$0 a translation vector. These parameters minimize a similarity metric (e.g., mutual information) plus regularization. This registration ensures anatomical consistency for tile extraction across subjects.

Tiles are defined in atlas space; to extract the corresponding subject patch for segmentation, the inverse transform is applied:

$k$1

This localization allows each network to process a spatially and anatomically consistent field across the dataset (Huo et al., 2018).

3. Network Architecture and Training

Each tile is segmented by an independent 3D U-Net, modified for $k$2 output classes. The canonical architecture is as follows:

• Input: $k$3 (single-channel, preprocessed intensity).
• Encoder: Four resolution levels; each contains two $k$4 convolutions with ReLU, followed by $k$5 max-pooling.
• Bottleneck: Two $k$6 convolutions + ReLU.
• Decoder: Four upsampling stages, each comprising a $k$7 transposed convolution (stride 2), skip connection from encoder, and two $k$8 convolutions with ReLU.
• Output: $k$9 convolution, 133 output maps, voxel-wise softmax.

Per-tile parameter count is approximately 30 million—comparable to a full 3D U-Net. All convolutions use zero-padding to preserve spatial extents. No geometric data augmentation is applied; however, intensity harmonization and N4 bias correction are performed during preprocessing (Huo et al., 2018, Huo et al., 2019).

4. Majority-Vote Label Fusion

After all $$\psi_n = [x_n : x_n + d_x,\; y_n : y_n + d_y,\; z_n : z_n + d_z],\;\; n=1,\dots,k.$$0 tile networks independently segment their respective atlas-aligned patches, their categorical outputs are combined in atlas space by voxel-wise majority voting:

$$\psi_n = [x_n : x_n + d_x,\; y_n : y_n + d_y,\; z_n : z_n + d_z],\;\; n=1,\dots,k.$$1

where $$\psi_n = [x_n : x_n + d_x,\; y_n : y_n + d_y,\; z_n : z_n + d_z],\;\; n=1,\dots,k.$$2 indicates whether tile $$\psi_n = [x_n : x_n + d_x,\; y_n : y_n + d_y,\; z_n : z_n + d_z],\;\; n=1,\dots,k.$$3 assigned label $$\psi_n = [x_n : x_n + d_x,\; y_n : y_n + d_y,\; z_n : z_n + d_z],\;\; n=1,\dots,k.$$4 to voxel $$\psi_n = [x_n : x_n + d_x,\; y_n : y_n + d_y,\; z_n : z_n + d_z],\;\; n=1,\dots,k.$$5. No explicit weighting per tile or probabilistic modeling of fusions is used in the baseline method. After fusion, the resulting segmentation can be warped to subject-native space using the inverse affine transform (Huo et al., 2018, Huo et al., 2019).

5. Auxiliary Label Generation for Pre-training

Given the scarcity of high-quality, manually traced training volumes (e.g., 45 OASIS/BrainCOLOR atlases), SLANT employs a two-stage training protocol:

• Stage 1: Pre-training on 5,111 T1-weighted MRI scans labeled via multi-atlas segmentation. For each, 133-label segmentations are created using a high-fidelity MAS pipeline (hierarchical non-local spatial STAPLE) with atlas selection, ANTs SyN registration, and local label fusion.
• Stage 2: Fine-tuning on the available manually labeled atlases ($$\psi_n = [x_n : x_n + d_x,\; y_n : y_n + d_y,\; z_n : z_n + d_z],\;\; n=1,\dots,k.$$6), adapting the pre-trained weights to the gold-standard annotations.

The per-tile loss function is voxel-wise cross-entropy

$$\psi_n = [x_n : x_n + d_x,\; y_n : y_n + d_y,\; z_n : z_n + d_z],\;\; n=1,\dots,k.$$7

and the optimizer is Adam with fixed learning rate $$\psi_n = [x_n : x_n + d_x,\; y_n : y_n + d_y,\; z_n : z_n + d_z],\;\; n=1,\dots,k.$$8, batch size 1 per GPU (Huo et al., 2018, Huo et al., 2019).

6. Empirical Performance and Computational Profile

SLANT-27 yields statistically significant improvements over MAS baselines for detailed (133-label) whole-brain segmentation across multiple cohorts. Representative mean Dice scores (mean ± std) are as follows (excluding background):

Method	OASIS (hold-out)	Colin27	CANDI (5–15 yr)
MAS (NLSS, 45 atlases)	0.760 ± 0.012	0.712	0.677 ± 0.029
SLANT-27	0.776 ± 0.012	0.732	0.711 ± 0.023

Key anatomical structures show superior Dice values for SLANT-27. For example, on the OASIS cohort: hippocampus (0.85 ± 0.02), thalamus proper (0.90 ± 0.01), and lateral ventricle (0.94 ± 0.01) (Huo et al., 2019).

Computation is efficient: inference requires ~15 minutes per scan on a single NVIDIA Titan Xp GPU (12 GB RAM), in contrast to >30 hours per scan for MAS baselines executed on a CPU cluster. The containerized SLANT pipeline ensures reproducible, end-to-end processing. The entire workflow is available as a Docker image and open-source codebase (see https://github.com/MASILab/SLANTbrainSeg) (Huo et al., 2018, Huo et al., 2019).

7. Limitations and Prospects

SLANT requires $$\psi_n = [x_n : x_n + d_x,\; y_n : y_n + d_y,\; z_n : z_n + d_z],\;\; n=1,\dots,k.$$9 network tiles; thus, training time and GPU requirements are proportional and can be resource-intensive absent multi-GPU hardware. The method relies on affine registration (and, for auxiliary labeling, ANTs SyN) which, while robust, adds preprocessing time (~5 min/scan). Explicit geometric augmentation was not used; consequently, generalizability to unseen contrasts or scanning protocols may be limited. Potential extensions include integration of more advanced 3D CNN modules (e.g., residual or attention mechanisms), learning-based registration, application to sub-millimeter or histological imaging, and development of improved fusion strategies (e.g., weighted voting, probabilistic modeling, or semi-supervised refinement) (Huo et al., 2019).

SLANT synthesizes traditional atlas-based image processing and modern deep learning, enabling detailed, high-accuracy, and computationally tractable whole-brain segmentation from limited labeled data by leveraging registration, tiling, and auxiliary pseudolabeling on a large scale.