AtlasSeg: Atlas Prior Guided Dual-U-Net for Cortical Segmentation in Fetal Brain MRI

TL;DR

This work tackles fetal brain MRI tissue segmentation across gestational ages by incorporating GA-specific atlas priors into a dual-U-Net framework. AtlasSeg processes MRI volumes and GA-matched atlas data in parallel, using a Multi-scale Attention Atlas Fusion Module to guide segmentation with anatomical priors. Across comparisons with six baselines, AtlasSeg achieves higher accuracy ($DSC$ up to $0.9105$) and greater robustness to contrast and noise, particularly at extreme GA ranges. The approach improves consistency across developmental stages and offers a practical tool for early prenatal diagnostics, with code released for public use.

Abstract

Accurate automatic tissue segmentation in fetal brain MRI is a crucial step in clinical diagnosis but remains challenging, particularly due to the dynamically changing anatomy and tissue contrast during fetal development. Existing segmentation networks can only implicitly learn age-related features, leading to a decline in accuracy at extreme early or late gestational ages (GAs). To improve segmentation performance throughout gestation, we introduce AtlasSeg, a dual-U-shape convolution network that explicitly integrates GA-specific information as guidance. By providing a publicly available fetal brain atlas with segmentation labels corresponding to relevant GAs, AtlasSeg effectively extracts age-specific patterns in the atlas branch and generates precise tissue segmentation in the segmentation branch. Multi-scale spatial attention feature fusions are constructed during both encoding and decoding stages to enhance feature flow and facilitate better information interactions between two branches. We compared AtlasSeg with six well-established networks in a seven-tissue segmentation task, achieving the highest average Dice similarity coefficient of 0.91. The improvement was particularly evident in extreme early or late GA cases, where training data was scare. Furthermore, AtlasSeg exhibited minimal performance degradation on low-quality images with contrast changes and noise, attributed to its anatomical shape priors. Overall, AtlasSeg demonstrated enhanced segmentation accuracy, better consistency across fetal ages, and robustness to perturbations, making it a powerful tool for reliable fetal brain MRI tissue segmentation, particularly suited for diagnostic assessments during early gestation.

Figures (5)

• Figure 1: (a) Example of the fetal brains, atlases, and corresponding tissue maps at younger, middle, and older ages. (b) Imbalanced distribution of fetal subjects and segmentation accuracy (yellow dots, from our previous work xu_site_2024) across GA, with reduced performance in the youngest and oldest fetal brains.
• Figure 2: (a) Architecture of AtlasSeg, which is built upon two parallel 3D U-Net backbones with encoder-decoder structures and skip connections. Dense multi-scale spatial attention fusion modules interlink two branches to build a dense feature flow. The network takes three 96×96×96 patches as input. (b) Convolutional blocks and operations used in (a). (c) Proposed MSA. This fusion module took image and atlas features from corresponding branches as inputs, and outputs fused feature for further segmentation. A group of convolutions with kernel sizes of [7,5,3,1] were used to extract multi-scale features.
• Figure 3: Segmentation performance across different fetal age groups using various networks in terms of DSC, 95HD, and ASSD. Bars indicate mean values, and error bars represent the SDs. The fetal age range from 22 to 38 weeks is divided into four groups, with each containing 4 or 5 ages. The asterisk labels the significant difference (*p$<$0.05, **p$<$0.01, ***p$<$0.001) with respect to AtlasSeg using paired t-test. The performance of AtlasSeg is the most accurate in each group and more robust across groups.
• Figure 4: Visual comparisons of segmentation labels predicted by different networks. (a) Multi-label predictions at the earliest (22 weeks) and latest (38 weeks) fetal ages. Incorrect predictions are highlighted with white arrows. (b) Single-label predictions, including the tissue of gray matter (GM, 23 weeks), white matter (WM, 37 weeks), and deep gray matter (dGM, 24 weeks). Correct, over, and under segmentations are indicated in blue, yellow, and green, respectively.
• Figure 5: Results of different networks on images with added perturbations. (a) and (b) involve gamma correction to adjust the image contrast, where image becomes brighter when $\gamma$$<$1 and darker when $\gamma$$>$1. (c) Gaussian noise with a standard deviation ranging from 0 to 0.05. AtlasSeg demonstrates superior robustness across all three tests. (d) Multi-label predictions of three perturbation tests on a 30-week fetal brain. The left three columns display the original image, the corrupted low-quality image, and the ground truth. The seven columns on the right show the segmentation labels generated by different networks. All six competing networks exhibit noticeable segmentation errors, while AtlasSeg maintains a significantly higher segmentation accuracy. SD = standard deviation.