FLAIRBrainSeg: Fine-grained brain segmentation using FLAIR MRI only

TL;DR

The paper tackles the challenge of fine-grained brain segmentation when T1-weighted MRI is unavailable by introducing FLAIRBrainSeg, a FLAIR-only method that yields 132-structure segmentations. It trains a 3D nnU-Net using supervision derived from T1w-based segmentations obtained after lesion inpainting, incorporating rigorous quality control and lesion-aware processing. The approach achieves high accuracy on in-domain and out-of-domain data and outperforms modality-agnostic baselines like SynthSeg and SynthSeg+ while requiring fewer computational resources. This modality-specific solution has practical clinical implications for MS and other conditions where T1w imaging is limited, with future work aimed at improving performance on low-resolution scans and broader lesion types, and plans for public release.

Abstract

This paper introduces a novel method for brain segmentation using only FLAIR MRIs, specifically targeting cases where access to other imaging modalities is limited. By leveraging existing automatic segmentation methods, we train a network to approximate segmentations, typically obtained from T1-weighted MRIs. Our method, called FLAIRBrainSeg, produces segmentations of 132 structures and is robust to multiple sclerosis lesions. Experiments on both in-domain and out-of-domain datasets demonstrate that our method outperforms modality-agnostic approaches based on image synthesis, the only currently available alternative for performing brain parcellation using FLAIR MRI alone. This technique holds promise for scenarios where T1-weighted MRIs are unavailable and offers a valuable alternative for clinicians and researchers in need of reliable anatomical segmentation.

Figures (6)

• Figure 1: Schematic overview of the T1w-based segmentation construction. From left to right, the images are : the (1) T1w and (2) FLAIR images, the (3) lesion maps, (4) inpainted T1w and (5) the obtained fined-grained segmentation.
• Figure 2: Left: Illustration of a synthetic FLAIR-like image created from T1w-based segmentation as part of the quality control process to address registration discrepancies. Right: Distribution of correlation scores on the training set. The vertical dotted line corresponds to the selected threshold.
• Figure 3: Average DSC for Experiment 1 and Experiment 2. Top: results on the in-domain dataset (i.e., OFSEP), Bottom: Results on the out-of-domain dataset (i.e., UKB), Left: Comparison of SynthSeg-132 and FLAIRBrainSeg over the 132 structures of the Neuromorphometrics protocol (Experiment 1), Right: Comparison of SynthSeg-132, SynthSeg$^{+}$, and FLAIRBrainSeg on the 35 common structures between Freesurfer and Neuromorphometrics protocols (Experiment 2).
• Figure 4: Examples of segmentation of 132 structures obtained with FLAIRBrainSeg and SynthSeg-132 in Experiment 1 are shown on different slices and images, for the in-domain (top) and out-of-domain (bottom) testing sets. The images were selected from cases with median mean DSC scores for SynthSeg-132 to ensure a fair representation of its performance. The top row highlights the accuracy of FLAIRBrainSeg, even in the presence of MS lesions, which are incorrectly labeled as ventricle or cortex by SynthSeg. The bottom row illustrates the differences between the segmentations obtained and the ground-truth labels on another example, further emphasizing the advantages of FLAIRBrainSeg in handling modality-specific challenges.
• Figure 5: Examples of obtained segmentations with FLAIRBrainSeg, SynthSeg-132, and SynthSeg$^{+}$. Please note that the SynthSeg$^{+}$ segmentation includes only the 35 selected structures for comparison which do not include cortical structures and cerebellum grey matter.