ABDSynth: Synthetic MRI Segmentation from CT

• ABDSynth is a method for abdominal MRI segmentation that uses CT segmentation maps to generate synthetic MRI images, addressing annotation scarcity.
• It employs a segmentation-conditioned Gaussian mixture model with extensive domain randomization to replicate diverse MRI contrasts and appearances.
• The system offers faster inference and lower computational cost while achieving competitive performance compared to MRI-trained models, despite slightly reduced accuracy.

ABDSynth is a SynthSeg-based model for generic abdominal MRI segmentation that is trained purely from CT segmentations and uses no real MRI images during training. It was introduced as a response to the central bottleneck in abdominal MRI segmentation: high inter-sequence and inter-scanner intensity variability combined with the cost of manual MRI annotation. In the benchmark that introduced it, ABDSynth is positioned as a sequence-agnostic and out-of-the-box alternative to MRI-trained systems, trading some accuracy for substantially reduced annotation requirements and lower computational cost (Krishnaswamy et al., 23 Jul 2025).

1. Definition and motivation

ABDSynth was proposed to answer a practical question: whether a useful abdominal MRI segmenter can be trained without any real MRI training images, using only widely available CT segmentations. The model is explicitly intended for multiple organ classes, generic abdominal MRI segmentation, and deployment scenarios in which annotated MRI data are scarce or expensive to obtain (Krishnaswamy et al., 23 Jul 2025).

The motivation follows directly from the characteristics of abdominal MRI. The benchmark paper identifies four related difficulties: MRI intensities vary substantially across scanners, sequences, and sites; MRI lacks the inherent intensity normalization available in CT; manual contouring in MRI is slow and expensive; and existing strong systems rely on iterative expert annotation loops. ABDSynth therefore reframes the training problem: instead of collecting real MRI images with labels, it synthesizes MRI-like images from anatomical segmentations and trains a supervised segmenter on those synthetic image-label pairs.

A frequent misconception is that ABDSynth is an MRI-trained model with synthetic augmentation. It is not. The paper states that ABDSynth is trained on CT segmentations only, with no real MRI images and no manually annotated MRI training set. Its purpose is precisely to reduce dependence on MRI labels rather than to complement them.

2. Synthetic training formulation

ABDSynth extends the SynthSeg idea to abdominal MRI by generating synthetic images from segmentation maps rather than learning directly from real intensity volumes. The high-level workflow is given in three steps: sample a CT segmentation from the training set, generate a synthetic volume using a segmentation-conditioned Gaussian mixture model (GMM) with randomized parameters, and use the resulting synthetic image-label pair to train a supervised 3D U-Net (Krishnaswamy et al., 23 Jul 2025).

The conditional image generation model is summarized as

$$p(x \mid y) = \sum_{k=1}^{K} \pi_k \,\mathcal{N}(x \mid \mu_k, \sigma_k^2),$$

where $x$ is the image intensity, $y$ is the label map conditioning the generation, $K$ is the number of Gaussian components, and $\pi_k$, $\mu_k$, and $\sigma_k^2$ are the mixture weights, means, and variances. The essential methodological point is that the image is generated from the label map, not inferred from paired MRI data.

To make the resulting segmenter robust to MRI heterogeneity, ABDSynth uses domain randomization. The generation parameters are sampled from very wide uniform distributions, so the synthetic volumes span many possible contrasts and appearances. The model further uses affine transforms, non-linear spatial warps, bias field corruption, contrast augmentation, noise injection, and simulated voxel resolution changes. The paper presents these choices as the mechanism by which the segmenter becomes robust to scanner differences, sequence differences, anisotropic resolution, field-of-view variation, and acquisition artifacts.

The training corpus consists of 128 segmentation maps from the TotalSegmentator CT training set. ABDSynth segments 33 regions, which is fewer than the other MRI systems in the benchmark: MRSegmentator segments 40 regions, MRISegmentator-Abdomen segments 62 regions, and TotalSegmentator MRI segments 59 regions.

3. Preprocessing, label subdivision, and optimization

The benchmark paper describes several refinements intended to enrich synthetic appearance and anatomical variability (Krishnaswamy et al., 23 Jul 2025). A single Gaussian per anatomical label is considered too coarse, so ABDSynth subdivides labels into finer subregions using expectation-maximization (EM) clustering). For foreground labels, the number of clusters is randomly chosen from

$K_{\text{FG}} \in \{1,2,3\},$

whereas for background it is randomly chosen from

$K_{\text{BG}} \in \{3,4,5,6,7\}.$

The paper states that the background receives more clusters because it is more heterogeneous.

The synthetic training setup also includes an MRI-specific pose prior. Arms are removed with probability 0.5 to simulate common trunk-only MRI acquisitions. The CT label maps are cropped or padded to $300 \times 300 \times 250$ at 1.5 mm isotropic resolution.

ABDSynth uses the same architecture and generation parameters as SynthSeg, trains for 500,000 iterations, and is optimized with a soft Dice loss. The paper gives the standard form

$x$0

where $x$1 are predicted soft probabilities, $x$2 are ground-truth labels, and $x$3 is a smoothing constant. Training takes about two weeks on an Nvidia A100 40GB GPU.

These details matter because ABDSynth is not merely a synthetic-data benchmark baseline. It is a fully specified segmentation pipeline whose performance depends on the interplay between label-conditioned generation, domain randomization, EM-based label subdivision, and large-scale augmentation.

4. Benchmark setting and empirical behavior

ABDSynth was evaluated on three public MRI datasets that were unseen during training for all compared methods (Krishnaswamy et al., 23 Jul 2025). The datasets span three manufacturers, five MRI sequences, healthy and diseased populations, and a wide range of resolutions and fields of view.

The evaluation datasets are as follows. AMOS MRI contains 60 subjects, comes from Philips scanners, includes diseased patients, and provides annotations for 15 abdominal organs. CHAOS contains 60 volumes total, with 20 per sequence, and includes T1 dual in-phase, T1 dual out-phase, and T2 SPIR scans from healthy subjects; its annotations cover the liver, spleen, and both kidneys. LiverHCCSeg contains 17 subjects with hepatocellular carcinoma and provides manual liver masks from two independent raters on T1 arterial phase MRI.

The benchmark uses Dice score for overlap and HD95 for boundary robustness. Its overall ranking is explicit:

1. MRSegmentator
2. TotalSegmentator MRI
3. ABDSynth
4. MRISegmentator-Abdomen

ABDSynth is described as usually slightly less accurate than the best MRI-trained systems, especially for smaller organs, variable morphology structures, organs with poor contrast, and sequence types it never saw in any real-image sense. At the same time, it is often competitive for the liver, spleen, and kidneys.

Dataset-specific observations sharpen that picture. On AMOS, all methods perform fairly well on liver, spleen, and kidneys, while ABDSynth is weaker on the stomach, gallbladder, and duodenum. On CHAOS, MRSegmentator is best overall, TotalSegmentator MRI and ABDSynth are described as respectable, and MRISegmentator-Abdomen performs worst because the sequences differ from its training set. On LiverHCCSeg, all methods are close in Dice, around 0.9–0.93, although HD95 reveals more variability.

A central empirical point is that ABDSynth never saw real MRI images during training, yet still segments real MRI scans across multiple sequences. The paper treats this as evidence that segmentation-conditioned synthetic training can generalize meaningfully across MRI domains, while also noting failures due to the reality gap between synthetic and real MRI appearance.

5. Computational profile and practical role

ABDSynth is the fastest and smallest model among the four systems compared in the benchmark (Krishnaswamy et al., 23 Jul 2025). The paper attributes this partly to the fact that ABDSynth does not use patch-based sliding-window inference and has about two-thirds fewer parameters than the MRI-trained baselines.

Method	Inference time (s)	Trainable parameters
MRSegmentator	57.95 ± 53.15	31M
MRISegmentator-Abdomen	98.19 ± 69.83	31M
TotalSegmentator MRI	39.60 ± 10.34	31M
ABDSynth	21.17 ± 19.30	13M

This computational profile defines ABDSynth’s practical niche. The benchmark recommends it when annotated MRI training data are scarce, when a method trained from widely available CT segmentations is preferred, when an out-of-the-box abdominal MRI segmenter is needed without sequence-specific retraining, when speed and lower model size matter, and when slightly lower accuracy is acceptable in exchange for much lower annotation burden.

The implementation is also intended to be usable rather than purely demonstrative. The paper states that the evaluation code and datasets are available through AbdoBench, and that inference code and weights for ABDSynth are provided there as well.

6. Limitations and significance

ABDSynth’s limitations are stated directly in the benchmark paper (Krishnaswamy et al., 23 Jul 2025). First, there is a reality gap: synthetic MRI does not perfectly match real MRI intensity distributions. Second, ABDSynth is generally slightly worse than MRSegmentator in accuracy. Third, it can suffer occasional outright failures on some subjects and sequences. Fourth, it predicts only 33 regions, which is fewer than the strongest competitors. The model is also reported to underperform on the stomach, gallbladder, pancreas, some right kidney cases, and some T1 out-phase CHAOS scans.

The study itself also has broader limitations. Not all available models were benchmarked, pathology-specific analysis was limited, some datasets lack pathology labels, annotation conventions differ across datasets, and specialized models such as TotalVibeSegmentator were not included.

Within those limits, ABDSynth functions as a proof-of-concept for a specific methodological claim: a domain-randomized, synthetic-data-trained segmenter can be useful for real abdominal MRI even when trained without any real MRI images. The benchmark’s final characterization is therefore deliberately balanced. ABDSynth is less accurate than the strongest MRI-trained models, but it is also much cheaper to train, faster to run, and more practical when MRI annotations are limited. In that sense, its importance lies less in surpassing MRI-trained segmentation systems than in establishing a viable alternative training paradigm for abdominal MRI segmentation under constrained annotation budgets.