Source-linked AI summary
Simultaneous synthesis of FLAIR and segmentation of white matter hypointensities from T1 MRIs
Mauricio Orbes-Arteaga, M. Jorge Cardoso, Lauge Sørensen, Marc Modat, Sébastien Ourselin, Mads Nielsen, Akshay Pai
TL;DR
Many datasets lack the FLAIR sequences used alongside T1 for WMH segmentation, while existing imputation and segmentation methods optimize the tasks separately. This paper jointly trains synthesis and segmentation; it reports nearly 3% mean Dice improvement over a baseline without imputation, along with better synthetic FLAIR images.
Problem
Many datasets lack desired MRI modalities such as FLAIR, although complementary modalities are commonly used for WMH segmentation.
Source: §1 · p. 2Method
Results
Nearly 3% mean Dice improvement over the baseline without imputation; the method also improved segmentation over offline synthesis and synthetic FLAIR quality over generation-only networks.
Source: §3.2 · p. 4, §4 · p. 7Takeaways & Limitations
Joint optimization reduced false positives and better captured small lesions, improving overall segmentation performance.
Source: §4 · pp. 6–7L2 reconstruction loss can blur images; adversarial alternatives add a network and may overfit with limited training data.
Source: §4 · p. 7
Abstract
from arXiv · showhide
Segmenting vascular pathologies such as white matter lesions in Brain magnetic resonance images (MRIs) require acquisition of multiple sequences such as T1-weighted (T1-w) --on which lesions appear hypointense-- and fluid attenuated inversion recovery (FLAIR) sequence --where lesions appear hyperintense--. However, most of the existing retrospective datasets do not consist of FLAIR sequences. Existing missing modality imputation methods separate the process of imputation, and the process of segmentation. In this paper, we propose a method to link both modality imputation and segmentation using convolutional neural networks. We show that by jointly optimizing the imputation network and the segmentation network, the method not only produces more realistic synthetic FLAIR images from T1-w images, but also improves the segmentation of WMH from T1-w images only.
1 Introduction
WMH segmentation matters for studying vascular pathology, but usually benefits from complementary MRI modalities that many datasets lack. The paper jointly synthesizes missing FLAIR images and segments WMH, reporting improved synthetic images and segmentation over preprocessing-based synthesis.
1 Introduction
- Manual WMH segmentation is impractical because of time demands and inter- and intra-rater variability.The paper links WMH segmentation to understanding vascular pathology in neurological disorders [Caligiuri et al. 2015].Source: §1 · pp. 1–2
- Most datasets contain only T1 or T1/T2/PD scans, limiting the available data with desired modalities for WMH segmentation.FLAIR is often acquired alongside T1 because the sequences provide complementary information.Source: §1 · p. 2
- Existing synthesis methods often use imputation as preprocessing, which can produce features that are not useful for final segmentation.CNN-based missing-modality methods include GAN approaches; the cited examples generate CT from MRI [Nie et al. 2017].Source: §1 · p. 2
- Jointly synthesizing FLAIR and segmenting WMH improved synthetic-image quality and T1-only segmentation over treating synthesis as preprocessing.The experiments used the 2017 WMH Segmentation Challenge dataset.Source: §1 · p. 2
2 Methods
The method addresses missing modalities by learning synthesis and segmentation together rather than optimizing them separately. A generator and classifier are trained iteratively, with a reconstruction loss and CNN-based network designs.
2 Methods
- When a modality is missing, a CNN generator G maps the available image Xa toward the missing image Xb for use with the classifier.The conventional imputation scheme separately synthesizes the missing modality and then uses it with the available modality for segmentation.Source: §2 · p. 2
- The generator and classifier are optimized simultaneously so classifier learning can reinforce synthesis of realistic images with features relevant to segmentation.The two networks are trained end to end and iteratively; the classifier uses real T1 and generated images to produce a WMH segmentation.Source: §2 · p. 3
- The generator uses an L2 reconstruction error between the real missing-modality image and its generated counterpart.The classifier can be viewed as a regularization term for the generator, or vice versa.Source: §2 · p. 3
2.1 Network architectures
- The segmentation network uses a U-Net, while the generation network modifies it for one T1 input channel and a LeakyReLU final layer.Both networks use Adam with learning rate 0.0002 and batch normalization, are trained iteratively at the same frequency, and use no data augmentation.Source: §2.1 · p. 3
3 Experiments and results
Experiments on the 2017 WMH Segmentation Challenge dataset compare proposed joint synthesis, offline synthesis, and T1-only segmentation. The proposed method improves segmentation and FLAIR reconstruction over offline synthesis, with qualitative results highlighting both its false-positive reduction and reconstruction trade-offs.
3.1 Data and Experiment
- The 60-subject dataset spans three clinics; six-fold cross-validation evaluates DSC, FPR, and FNR.Each fold uses 10 test subjects, 5 for validation, and 45 for training.Source: §3.1 · p. 4
3.2 Results
- Nearly 3% mean Dice improvement over the baseline without imputation; the proposed method also improves segmentation over offline synthesis.Table 1 reports average performance measures for all methods.Source: Table 1 · p. 4, §3.2 · p. 4
- Proposed-method FPR is 20.47% lower than Unimodal and 5.3% lower than offline synthesis, while Unimodal has the lowest FNR.The authors report the FPR reductions as evidence of fewer false positives.Source: §3.2 · p. 4
- Visual comparisons show fewer false positives with the proposed method than Unimodal, while synthesis methods tend to share false negatives that may reflect image blurring.Unimodal false positives include small cortical regions and lesion-border regions; both synthesis methods perform well on large contiguous lesions, but poorly on small diffuse WMHs.Source: §3.2 · pp. 5–6, Table 2 · p. 5
3.3 Results of Generation
- Generated FLAIR achieves average PSNR 11.01 versus 9.65 offline, and MAE 0.26 versus 0.31 offline.The authors report better reconstruction by both measures for the proposed method.Source: §3.3 · p. 6, Table 3 · p. 6
- Both generators respond similarly to low lesion loads and well to large contiguous lesions, but outputs are blurred; the proposed method is more sensitive to small diffuse WMHs.The authors describe L2-based optimization as balancing structural preservation against blur.Source: §3.3 · p. 6
4 Discussion and Concluding Remarks
Joint optimization uses FLAIR to guide synthesis and segmentation, improving segmentation from T1-only images while producing better synthetic FLAIR. The method captures small lesions, though its L2 loss can blur images.
4 Discussion and Concluding Remarks
- Independent synthesis tends to under-segment, whereas the joint strategy captures small lesions and improves overall segmentation.The passage also describes excessive over-segmentation in T1-based segmentation and fewer false positives when using FLAIR information through a generator.Source: §4 · pp. 6–7
- Joint optimization better adapts to capture small lesions, leading to significantly better overall segmentation performance.FLAIR sequences drive optimization, and the method reduces false positives associated with T1-based over-segmentation.Source: §4 · pp. 6–7
- The proposed method produces better synthetic FLAIR images than generation-only networks, with lesions barely visible in T1 appearing in the synthetic images.The authors suggest complementary information from joint optimization with the segmentation network may explain the improvement.Source: §4 · p. 7
- L2 loss can blur synthetic images; adversarial loss may address this, but an additional network and limited training data may increase overfitting risk.The authors present L2 as a compromise between model complexity and segmentation performance.Source: §4 · p. 7