Source-linked AI summary
Automatic brain tissue segmentation in fetal MRI using convolutional neural networks
N. Khalili, N. Lessmann, E. Turk, N. Claessens, R. de Heus, T. Kolk, M. A. Viergever, M. J. N. L. Benders, I. Isgum
TL;DR
Manual fetal brain tissue segmentation is time-consuming and challenged by intensity inhomogeneity in fetal MRI. The paper uses two CNNs and synthetic-artifact augmentation to segment seven tissue classes; the authors report improved segmentation metrics with augmentation. Severe-motion slices were excluded from training and testing.
Problem
Manual fetal brain tissue segmentation is time-consuming, and intensity inhomogeneity from fetal motion complicates automatic segmentation.
Source: §1 · pp. 2, 5Method
Two CNNs extract intracranial volume and segment seven tissue classes, with training slices augmented by synthesized intensity inhomogeneity.
Source: §1 · p. 6, §3.2 · pp. 11–12Results
The authors report that simulated-artifact augmentation improved segmentation performance and enabled accurate segmentation when training references came only from artifact-free slices.
Source: §6 · pp. 26–27Takeaways & Limitations
Synthetic intensity-inhomogeneity augmentation may complement or replace preprocessing such as bias-field correction, while allowing training from easier-to-annotate artifact-free data.
Source: §6 · p. 26Slices with severe motion artifacts were excluded from training and testing because manual annotation for evaluation was hardly feasible.
Source: §6 · p. 28
Abstract
from arXiv · showhide
MR images of fetuses allow clinicians to detect brain abnormalities in an early stage of development. The cornerstone of volumetric and morphologic analysis in fetal MRI is segmentation of the fetal brain into different tissue classes. Manual segmentation is cumbersome and time consuming, hence automatic segmentation could substantially simplify the procedure. However, automatic brain tissue segmentation in these scans is challenging owing to artifacts including intensity inhomogeneity, caused in particular by spontaneous fetal movements during the scan. Unlike methods that estimate the bias field to remove intensity inhomogeneity as a preprocessing step to segmentation, we propose to perform segmentation using a convolutional neural network that exploits images with synthetically introduced intensity inhomogeneity as data augmentation. The method first uses a CNN to extract the intracranial volume. Thereafter, another CNN with the same architecture is employed to segment the extracted volume into seven brain tissue classes: cerebellum, basal ganglia and thalami, ventricular cerebrospinal fluid, white matter, brain stem, cortical gray matter and extracerebral cerebrospinal fluid. To make the method applicable to slices showing intensity inhomogeneity artifacts, the training data was augmented by applying a combination of linear gradients with random offsets and orientations to image slices without artifacts.
1. Introduction
Fetal MRI tissue segmentation is clinically useful but manual labeling is time-consuming, expertise-intensive, and complicated by motion-related image artifacts. This paper proposes CNN-based segmentation with synthetic intensity-inhomogeneity augmentation and seven tissue classes.
1. Introduction
- Manual fetal brain segmentation is time-consuming and requires expertise, while fetal motion can degrade image quality and cause intensity inhomogeneity.Fetal imaging also has limited image quality because receiver coils cannot be placed close to the anatomy of interest.Source: §1 · p. 2
- Synthetic intensity-inhomogeneity augmentation is proposed to help CNNs adapt to artifacts without requiring complex preprocessing at inference.The synthetic changes alter image intensities, not structure orientation or shape, so existing reference segmentations remain usable.Source: §1 · p. 5
- Cerebellum and brain-stem characterization matters alongside the main tissue classes because these regions inform understanding of healthy or aberrant development.The cerebellum is among the fastest-growing brain regions during the last trimester.Source: §1 · p. 5
- Fetal scans have a large field of view that includes maternal and fetal body regions, motivating automatic intracranial-volume extraction to define the region of interest.The proposed extraction operates on the full image before cropping to the brain.Source: §1 · pp. 5–6
- The method augments training slices with synthetic intensity inhomogeneity and segments fetal brain tissue into seven classes, beyond the three-class focus of previous methods.It first identifies the intracranial volume, then segments the extracted volume.Source: §1 · p. 6
2. Data
The study evaluates fetal MRI segmentation using coronal scans from 12 fetuses and also examines applicability to neonatal MRI. Fetal annotations provide seven tissue labels, while inhomogeneous neonatal test scans lack manual references.
2.1. Fetal MRI dataset
- The fetal dataset comprised 12 T2-weighted scans acquired in the coronal plane, with images reconstructed without the high-resolution framework.The fetuses were 22.9–34.6 weeks post menstrual age; the images were acquired at 3T.Source: §2.1 · pp. 6–7
- Manual fetal reference annotations labeled seven tissue classes in 2D slices, with intracranial volume defined as their union.Fifteen severely motion-distorted slices could not be annotated; 126 annotated slices were identified as affected by intensity inhomogeneity.Source: §2.1 · p. 7
2.2. Neonatal MRI dataset
- Nine preterm-infant brain MR scans were included to assess whether the intensity-inhomogeneity augmentation might apply beyond fetal MRI.Four scans without visible inhomogeneity had expert annotations; four inhomogeneous scans formed a test set evaluated only by visual inspection because manual references were unavailable.Source: §2.2 · pp. 7–8
3. Method
The method first extracts fetal intracranial volume and then segments it into seven tissue classes using separately trained 2D CNNs. Training augmentation simulates intensity inhomogeneity with randomized gradients while retaining the original labels.
3. Method
- The pipeline extracts the fetal intracranial volume before segmenting it into seven tissue classes, using the same network architecture for both stages.The identified volume is cropped to the region of interest before tissue segmentation.Source: §3 · p. 8, §3.1 · p. 10
3.1. Brain segmentation
- Two independently trained fully convolutional networks use identical U-net architectures on 2D slices for intracranial-volume extraction and tissue segmentation.The tissue network classifies pixels into seven tissues or background and maximizes Dice to address class imbalance.Source: §3.1 · pp. 8, 11
- The intracranial-volume network minimizes cross-entropy and removes small isolated 3D components below an empirically chosen 3 cm^3 threshold.The threshold is intended to remove false-positive clusters without removing brain portions disconnected by motion-related signal loss.Source: §3.1 · pp. 10–11
3.2. Intensity inhomogeneity augmentation (IIA)
- Intensity-inhomogeneity augmentation applies combinations of linear gradients with randomized offsets and orientations to artifact-free slices.Random offsets and rotation vary the location and orientation of low-contrast regions while leaving the reference segmentation applicable.Source: §3.2 · pp. 11–12
4. Evaluation
Evaluation used Dice coefficient for volume overlap and mean surface distance between manual references and automatic segmentations.
4. Evaluation
- Evaluation used Dice coefficient for volume overlap and mean surface distance between manual references and automatic segmentations.Metrics were computed per slice and averaged across fetal scans, but in 3D for neonatal scans.Source: §4 · p. 12
5. Experiments and Results
Experiments evaluated fetal segmentation across tissue classes, augmentation strategies, artifact levels, and neonatal scans. The proposed method was comparable to earlier methods despite finer tissue classification, while the cross-method comparison is indicative only because datasets differed.
5.1. Segmentation performance
- Across six fetal test scans, average Dice coefficients ranged from 0.80 for CB to 0.94 for eCSF, while mean surface distances ranged from 0.62 mm for CB to 0.18 mm for BS.Slices with intensity inhomogeneity strong enough to hamper manual annotation were less accurately segmented than slices without visible inhomogeneity.Source: §5.1 · pp. 14–15, Table 2 · p. 16
- The proposed method performed comparably to earlier methods while segmenting seven tissue classes rather than three or four, but different datasets make this comparison indicative only.The previous methods' results were taken from the literature.Source: §5.1 · p. 15, Table 4 · p. 20
5.2. Comparison of data augmentation techniques
- With IIA, performance improved across scans, differences between scans narrowed, and the heavily artifact-affected scan reached accuracy comparable to scans with fewer artifacts.Standard augmentation alone did not achieve the same reduction in between-scan performance differences.Source: §5.2 · pp. 19–20, Figure 9 · p. 21
- All augmentation methods improved segmentation, and intensity inhomogeneity augmentation further improved all tissue classes; CB Dice coefficient rose from 0.3 to 0.8 across the augmentation comparisons.Flipping and rotation noticeably improved CB, BS, vCSF, and BGT, while IIA also improved WM, cGM, and eCSF.Source: §5.2 · p. 20, Figure 10 · p. 22
5.3. Impact of extent of intensity inhomogeneity augmentation
- Using simulated intensity inhomogeneity on 20% of training slices already improved performance substantially, with comparable performance across 20% to 100%.Training exclusively on simulated slices performed slightly worse than mixing original and manipulated slices.Source: §5.3 · pp. 20, 22, Figure 11 · p. 23
5.4. Evaluation of IIA on neonatal brain segmentation
- In neonatal scans with visible artifacts but no manual reference segmentations, visual inspection found more accurate results with IIA, particularly for BGT, BS, vCSF, and eCSF.The two test scans with reference segmentations showed no visible intensity inhomogeneity artifacts.Source: §5.4 · p. 24, Table 5 · p. 24
6. Discussion
The authors report state-of-the-art fetal brain segmentation using fewer training scans, with simulated intensity inhomogeneity augmentation improving performance across tissue classes. The discussion also identifies evaluation and training-scope boundaries.
6. Discussion
- The pipeline uses two identical U-net CNNs: one extracts intracranial volume, and the other segments it into seven tissue classes.Training with simulated artifacts supported accurate segmentation from manual references on artifact-free slices, addressing the scarcity of annotations for compromised data.Source: §6 · pp. 24, 26
- With six fetal scans, the method achieved state-of-the-art segmentation performance compared with prior methods using 20 and up to 80 scans.The authors note that manual annotation of a seven-class fetal brain scan takes about 40 hours, so fewer training scans reduce annotation effort and costs.Source: §6 · p. 26
- Simulated intensity inhomogeneity may improve robustness and potentially replace or complement preprocessing such as bias-field correction or volumetric reconstruction.The authors argue that artifact-free slices are easier and more accurate to annotate than compromised data.Source: §6 · p. 26
- Random flipping and rotation improved cerebellum, brain stem, basal ganglia and thalami, and ventricular CSF, while adding simulated inhomogeneity improved performance across scans.Without inhomogeneity augmentation, Dice improved in all but one scan; that scan had intensity inhomogeneity in nearly all slices.Source: §6 · pp. 26–27
- Simulated intensity inhomogeneity further improved segmentation across all tissue classes, including low-contrast white matter and cortical gray matter.It helped overcome frequent reciprocal confusion between ventricular and extracerebral CSF, possibly by encouraging use of shape and context as well as intensity.Source: §6 · pp. 26–27
- Evaluation used coronal fetal MRI; applying the supervised method to axial or sagittal scans requires manually annotated training data.The authors relied on U-net and did not evaluate IIA with other architectures, although they expect it may help other supervised networks.Source: §6 · p. 27
- In neonatal MRI, visual inspection showed substantial improvement on artifact-affected slices, while artifact-free images showed slight quantitative improvement.Performance remained below the authors’ earlier method because this network was not specifically adjusted for neonatal tissue segmentation; Figure 12 contrasts outputs with and without IIA.Source: §6 · pp. 25, 28, Figure 12 · p. 25
- The study excluded a few slices with severe motion artifacts, and generating such slices for future work remains an open direction.Training used representative data; transfer learning with non-representative data was proposed as a possible way to increase the training sample size and potentially improve performance.Source: §6 · p. 28
7. Conclusion
The paper presents a CNN method for segmenting fetal MRI into seven tissue classes, using synthesized intensity-inhomogeneity artifacts in training to help the network cope with them. This approach could potentially complement or replace preprocessing such as bias-field correction and improve segmentation.
7. Conclusion
- The method uses convolutional neural networks to segment fetal MRI into seven brain tissue classes.Source: §7 · p. 28
- Training data augmented with synthesized intensity-inhomogeneity artifacts helps the network cope with those artifacts.Source: §7 · p. 28
- The approach could potentially complement or replace preprocessing such as bias-field correction and improve segmentation.Source: §7 · p. 28