Patch-wise 3D U-Net segments six-month-old infant brain MRI in benchmark study

A patch-wise three-dimensional U-Net segmented gray matter, white matter and cerebrospinal fluid in MRI scans from six-month-old infants, achieving Dice similarity scores of 0.92, 0.80 and 0.74, respectively, in a benchmark study using the iSeg-2017 dataset. The model also recorded 95th-percentile Hausdorff distances of 1.0 mm, 1.4 mm and 2.1 mm, and average surface distances of 0.25 mm, 0.36 mm and 0.48 mm for the three tissues.

The research focused on the isointense stage of infant brain development, when gray and white matter have similar intensities in both T1- and T2-weighted MRI. This low contrast, together with myelination, noise and motion-related errors, makes automated tissue separation difficult. The authors said the benchmark was particularly challenging because ongoing myelination causes substantial overlap between the intensity distributions of gray and white matter.

The study proposed a 3D patch-wise U-Net with weighted-average reconstruction. Each MRI volume was divided into fixed-size 64 × 64 × 64-voxel patches with a stride of 32 voxels along each dimension, creating 50% overlap between adjacent patches. Patch-level segmentation reduced the memory required to process a complete high-resolution volume while retaining local anatomical information.

The pipeline comprised MRI loading and depth-wise cropping, patch creation, intensity normalization and label mapping, patch-wise segmentation, and reconstruction of the complete 3D volume. Normalized image intensities were scaled to the range from zero to one by dividing each patch by its maximum voxel value. Labels were remapped to consecutive indices for cerebrospinal fluid, gray matter and white matter.

The U-Net had a four-level encoder-decoder structure with skip connections. Its encoder used paired 3 × 3 × 3 convolutions, batch normalization and ReLU activation, followed by 2 × 2 × 2 max pooling. Feature channels increased from 8 to 64 across the encoder, while a two-layer bottleneck operated at the lowest resolution. The decoder used 3D transposed convolutions to restore spatial resolution and concatenated encoder features through skip connections. A final 1 × 1 × 1 softmax layer produced probabilities for background, gray matter, white matter and cerebrospinal fluid.

After separate patch predictions were made, overlapping regions were combined with spatial weights based on each voxel’s distance from the center of the patches covering it. The paper states that voxels near patch centers received greater weight, while voxels near patch boundaries received less, producing smoother transitions and reducing stitching artifacts. This reconstruction method was designed to address boundary discontinuities that can arise when separately processed patches are joined.

Training and evaluation used co-registered T1- and T2-weighted MRI from six-month-old infants acquired on a Siemens 3T scanner at 1 mm isotropic resolution. The dataset contains 10 training participants and 13 official test subjects, with expert manual segmentations of the three tissues. The training participants were divided subject-wise into nine for training and one for validation, keeping all patches from each participant within a single set. Although both modalities were used in preprocessing, segmentation was conducted primarily on T1-weighted images.

The model was trained for 1,000 epochs with a combined Dice and cross-entropy loss, the AdamW optimizer and a cosine-annealed learning rate. Training used mixed-precision arithmetic on an NVIDIA GeForce RTX 4080 with 12 GB of memory. The reported training footprint was about 1.9 GB, rising to 4.2 GB when activations and gradients were included, and convergence took approximately 50 to 65 hours. Inference for a 256 × 256 × 192 volume took about 2.5 seconds, with another 0.5 seconds for patch reconstruction.

On the 13-subject test set, the model achieved 95% overall voxel-wise accuracy, a weighted-average F1 score of 0.95 and a macro-average F1 score of 0.87. Gray matter had precision of 0.93, recall of 0.90 and F1 of 0.92; white matter had precision of 0.80, recall of 0.82 and F1 of 0.81; and cerebrospinal fluid had precision of 0.75, recall of 0.73 and F1 of 0.74. The reported 95% confidence ranges for the tissue-specific F1 scores did not overlap, and the paper reported statistically significant differences for all pairwise comparisons at p < 0.05.

The results indicate a trade-off between volumetric overlap and boundary precision. Some comparison methods reported higher Dice scores for white matter and cerebrospinal fluid, but the proposed model produced smaller HD95 and average surface distances in the comparisons presented. Performance was weaker for cerebrospinal fluid, whose thin and irregular structures are affected by partial-volume effects and ambiguous boundaries. Generalizability remains untested beyond the benchmark’s single scanner, single age point and small cohort, with no direct validation in other age groups, scanners or multicenter data. The authors proposed future work on patch and stride sensitivity, interpretability, hybrid CNN-transformer models and workflow integration, but none of those extensions had been evaluated in this study.

Subject of Research: Automated segmentation of infant brain tissues in six-month-old MRI using a patch-wise 3D U-Net

Article Title: Deep learning model for infant brain tissue segmentation using 3D patch-wise U-Net with weighted-average reconstruction

Article References: Mhlanga, S., & Viriri, S. (2026). Deep learning model for infant brain tissue segmentation using 3D patch-wise U-Net with weighted-average reconstruction. Discover Artificial Intelligence, 6(1), Article 1277. https://doi.org/10.1007/s44163-026-01982-x

Image Credits: AI Generated

DOI: 10.1007/s44163-026-01982-x

Keywords: infant brain MRI, tissue segmentation, 3D U-Net, patch-wise learning, weighted reconstruction, iSeg-2017, gray matter, white matter, cerebrospinal fluid, boundary accuracy, Deep, learning