Deep infant brain segmentation from multi-contrast MRI

TL;DR

This work tackles the challenge of robust infant brain segmentation across diverse ages and MRI protocols by introducing BabySeg, a group-input segmentation framework that uses domain randomization and a data engine to synthesize varied contrasts. The core idea is a U-Net with group convolutions that can ingest a flexible number of input scans, enabling segmentation from any combination of modalities without retraining. The authors demonstrate state-of-the-art performance across multiple age cohorts and input configurations with significantly faster runtimes, while acknowledging limitations in finer cerebellar structures and the need for initial image alignment. Overall, BabySeg offers a scalable, versatile tool for pediatric neuroimaging that can facilitate large-scale analyses in research and clinical settings.

Abstract

Segmentation of magnetic resonance images (MRI) facilitates analysis of human brain development by delineating anatomical structures. However, in infants and young children, accurate segmentation is challenging due to development and imaging constraints. Pediatric brain MRI is notoriously difficult to acquire, with inconsistent availability of imaging modalities, substantial non-head anatomy in the field of view, and frequent motion artifacts. This has led to specialized segmentation models that are often limited to specific image types or narrow age groups, or that are fragile for more variable images such as those acquired clinically. We address this method fragmentation with BabySeg, a deep learning brain segmentation framework for infants and young children that supports diverse MRI protocols, including repeat scans and image types unavailable during training. Our approach builds on recent domain randomization techniques, which synthesize training images far beyond realistic bounds to promote dataset shift invariance. We also describe a mechanism that enables models to flexibly pool and interact features from any number of input scans. We demonstrate state-of-the-art performance that matches or exceeds the accuracy of several existing methods for various age cohorts and input configurations using a single model, in a fraction of the runtime required by many existing tools.

Figures (8)

• Figure 1: BabySeg framework. At every step, we sample images and a label map from a subject. First, we randomly move and deform these inputs, adding synthetic background structures to the label map. Second, we synthesize training images and apply a series of corruptions. Third, the BabySeg U-Net predicts a single segmentation map from these inputs.
• Figure 2: T1-weighted contrast. At birth, the cerebral cortex appears bright and the white matter dark. By year 1, this pattern reverses. During the inversion phase, overall contrast is low and may vary within a single scan. At month 4, for example, the shown subject has bright anterior and dark posterior cortex.
• Figure 3: Group convolution. The convolutional layer supports a flexible number of $n$ inputs by first separately convolving each group entry $\mathbf{f}^\mathrm{in}_i$, where $i \in \{1, 2, ..., n\}$, as well as convolving the mean group representation $\bar{\mathbf{f}}^\mathrm{in}$. It encourages interaction across group entries by averaging each with the convolved group mean to output $\mathbf{f}^\mathrm{out}_i$.
• Figure 4: Group segmentation network. BabySeg implements a U-Net, which pools features at multiple resolutions. Each gray layer applies a group convolution, interacting features from a flexible number of inputs. A group fusion layer averages across groups to output a single label map delineating $K$ anatomical structures including background. Numbers indicate filter count.
• Figure 5: Data engine. First, we sample, move, and warp images and the associated label map, adding synthetic background blobs to the latter. Second, we remap intensities and generate images from the label map. Third, we apply a series of corruptions, to obtain images for training. All steps are randomized.