Project

Liver Fat Fraction Assessment by Ultra-Low Dose CT

MRI-supervised 3D liver fat fraction mapping — quantifying fatty liver at the radiation dose of a two-view chest X-ray, making CT-based NAFLD screening of asymptomatic populations plausible for the first time.

Lilach Arbel Issaschar · Lawrence Carl · Prof. Nahum Kiryati · Dr. Michael Green · Prof. Edith Michelle Marom · Dr. Arnaldo Mayer · Computational Imaging Lab, Diagnostic Imaging, Sheba Medical Center · School of Electrical & Computer Engineering and Gray Faculty of Medicine, Tel Aviv University

Grid of four test cases. Each row shows the MRI-derived ground-truth fat fraction map, then predictions from the proposed NN+DN-80 model, NN-80, LR-80 (all from 80 kVp input) and NN-140, LR-140 (from 140 kVp input). The proposed method closely matches the ground truth while the linear regression maps are visibly noisy. — Predicted fat fraction maps for four external test cases. The proposed method (NN+DN-80) uses only ultra-low dose 80 kVp CT at inference, yet produces maps that closely track the MRI-derived ground truth — while the clinically standard linear regression (LR-80, LR-140) is visibly noisy and unreliable. *Source: MICCAI ML-CDS 2025 paper, Fig. 3.*

0.075–0.15 mSv

Effective dose — comparable to an AP/LAT chest X-ray (~0.1 mSv)

0.79

Cross-validated R² of the proposed model, up from 0.59 without denoising

80 kVp

Single-energy ultra-low dose CT is the only input needed at inference

Prospective subjects with same-day paired ULD-CT and MRI

Why fatty liver needs a screening test

Non-alcoholic fatty liver disease (NAFLD) — the accumulation of fat within liver cells — is the most prevalent chronic liver disease worldwide, affecting roughly 25% of the global population. About 5% of cases progress to non-alcoholic steatohepatitis (NASH), which raises the risk of fibrosis, cirrhosis and hepatocellular carcinoma, and is the second leading cause of liver transplantation in the United States.

Critically, NAFLD is reversible when caught early: diagnosis and care at an early stage can halt or reverse disease progression through lifestyle intervention or medical treatment. That makes NAFLD an unusually strong candidate for population screening — yet no screening test for it exists in practice today.

The gap in current diagnosis

Every existing method for liver fat quantification trades away something a screening test cannot afford to lose:

Method	Sensitivity at low fat %	3D map	Radiation	Accessibility
Ultrasound	Poor, operator dependent	✗	None	✓
CT	Low	✓	High	✓
MRI	Very high	✓	None	✗
Biopsy	Gold standard	✗	None	✗

Biopsy is the gold standard but is invasive and samples a tiny volume of the liver. MRI-derived fat fraction mapping is non-invasive and highly accurate, but costly and inaccessible at population scale. Ultrasound is cheap and available everywhere, but operator-dependent and unreliable precisely where screening matters most — at low fat fractions. CT is accessible and produces true 3D maps, but at normal dose its radiation burden rules it out for asymptomatic people.

There is currently no method that is simultaneously safe, accurate, accessible, and 3D — the four properties a fatty liver screening test would need.

Our goal: close the gap with ultra-low dose CT

CT already has three of the four properties. The problem is dose. So we attack dose directly: acquire the liver at an ultra-low dose, and recover MRI-grade fat fraction accuracy computationally — by training the model against MRI-derived fat fraction maps.

Horizontal bar chart on a logarithmic scale comparing effective dose: standard 120 kVp CT at 8.0 mSv, 140 kVp ULDCT at 0.17 mSv, an AP/LAT chest X-ray at 0.1 mSv, and 80 kVp ULDCT at 0.075 mSv. — Effective dose on a log scale. The 80 kVp ultra-low dose scan sits **below** a two-view chest radiograph, and roughly two orders of magnitude below a standard 120 kVp abdominal CT — a ~99% reduction.

The clinical method today relates unenhanced CT Hounsfield Units to fat fraction with a linear fit calibrated for 120 kVp scans. That relationship holds reasonably at normal dose and 120 kVp, but it degrades at other tube voltages, and at ultra-low doses the noise is so severe that linearity effectively breaks down. Recovering fat fraction from such data requires more than a line.

A unique prospective dataset

35 subjects were prospectively recruited at Sheba Medical Center and gave informed consent. Each received, on the same day: a pair of ultra-low dose CT scans at 80 kVp and 140 kVp on a single-energy Revolution CT scanner (GE Healthcare), and in-phase / opposed-phase gradient-echo MRI on a 3T Ingenia scanner (Philips). Tube current was fixed by BMI — 10/20 mA (140 kVp) and 20/40 mA (80 kVp) for BMI ≤ 29 and BMI > 29 respectively.

The MRI in-phase (IP) and opposed-phase (OP) images yield the ground-truth fat fraction map through FF = (IP − OP) / (2 · IP). Across the cohort, MRI-derived liver fat fraction spans 0.7% to 34.3% (median 4.5%), covering healthy livers through severe steatosis.

Five registered axial slices from one subject: CT at 80 kV, CT at 140 kV, MRI in-phase, MRI opposed-phase, and the derived fat fraction map shown as a colour overlay on the liver with a 0–40% scale. — One case from the dataset. Left to right: 80 kVp ULD-CT, 140 kVp ULD-CT, MRI in-phase, MRI opposed-phase, and the MRI-derived fat fraction map used as ground truth. All modalities are registered. Note how much noisier the 80 kVp scan is than the 140 kVp scan acquired moments later.

Before training, the liver is segmented in CT with a SwinUNETR model and in MRI with an Attention U-Net, and the MRI and 80 kVp scans are registered to the 140 kVp scan using SuperElastix in a two-step rigid then non-rigid process.

Method: joint denoising and fat fraction mapping

The framework has two neural blocks — a denoiser and a fat fraction regressor — trained with two different supervision signals, and then fine-tuned together end-to-end. Both supervision signals are used only at training time. At inference the network sees nothing but the single-energy 80 kVp scan.

The proposed framework. The denoising block is trained with 80 kVp input, 140 kVp ground truth and a VGG perceptual loss; the fat fraction regression block takes both the raw and denoised 80 kVp volumes and is trained against the MRI-derived fat fraction map with a masked L1 loss. Everything inside the dashed box runs at inference — the 140 kVp scan and the MRI are never needed once the model is trained. Redrawn after the paper, Fig. 1.

Ultra-low dose CT denoising

The 80 kVp scans are extremely noisy — too noisy, in their raw form, for accurate fat fraction estimation. We denoise them using the 140 kVp scan of the same subject as the training target. Although the 140 kVp scan is itself an ultra-low dose acquisition at the same fixed tube current, the higher tube voltage produces a markedly less noisy image, which makes it a usable supervision signal without ever exposing the patient to a normal-dose scan.

The denoiser is a convolution-only 2D DnCNN trained on random 64 × 64 patches with a VGG16 perceptual loss. Perceptual loss was chosen over pixel-wise L1 or MSE deliberately: we want the network to learn high-level structural and textural features rather than to reproduce the exact HU values of the 140 kVp scan, so that it enhances the 80 kVp input while preserving clinically relevant content.

Three axial abdominal CT slices side by side: the grainy 80 kVp input, the much smoother denoised output, and the 140 kVp target, which retains moderate noise. — Denoising block in action. Left: the raw 80 kVp ultra-low dose input. Centre: the denoised output. Right: the 140 kVp target used for supervision — itself an ultra-low dose scan, just less noisy.

Fat fraction regression

Regression is performed by a SwinUNETR operating on random 3D patches of 32 × 128 × 128, initialised from pre-trained Swin transformer weights. Both the original 80 kVp volume and its denoised version are fed to the network, giving it more to work with than a single-energy input alone. Training uses a masked, voxel-wise L1 loss computed only over liver voxels.

Four panels for one case: MRI ground-truth fat fraction overlay reading 9.21 ± 4.04, the predicted fat fraction from 80 kVp plus denoising reading 8.83 ± 3.94, and two CT scans in HU with the liver overlaid. — Predicted versus ground-truth fat fraction for one case. The model, given only the ultra-low dose 80 kVp scan, predicts a mean liver fat fraction of 8.83 ± 3.94% against an MRI-derived ground truth of 9.21 ± 4.04%.

Joint fine-tuning

After each network is trained individually, the two are fine-tuned jointly from their respective pre-trained weights, optimising only the masked L1 distance on the final fat fraction prediction. No regularisation is applied to the denoised intermediate output — the denoiser is free to adapt its parameters away from "looking like a 140 kVp scan" and toward whatever representation best serves the fat fraction estimate.

Experiments

The 35 cases were split into an external test set of 10 and a cross-validation pool of 25, stratified by fat fraction so both reflect the same distribution. The 25 were used for five-fold cross-validation — each fold of 5 serving as validation once, with the other 20 for training.

Five models were compared, at two energies:

LR-80 / LR-140 — linear regression from CT Hounsfield Units to fat fraction, fitted at each energy. This is the current clinical approach.
NN-80 / NN-140 — the neural fat fraction regressor alone, on raw input.
NN+DN-80 — the proposed model: denoiser and regressor jointly fine-tuned, on 80 kVp input only.

Fitting the linear baselines across folds gives FF [%] = −0.26 · HU + 14.53 at 80 kVp and FF [%] = −0.58 · HU + 34.58 at 140 kVp. All experiments ran on a single NVIDIA RTX A6000 (48 GB) in PyTorch and MONAI.

Results

Every neural model substantially outperforms its linear-regression baseline. Both linear baselines produce a negative R² on the external test set — meaning they predict fat fraction worse than simply guessing the cohort mean.

Model	Cross-validation L1 ↓	Cross-validation R² ↑	Test L1 ↓	Test R² ↑
80 kVp only
LR-80	4.75 ± 0.84	−0.37 ± 0.87	6.45 ± 0.29	−0.39 ± 0.16
NN-80	3.20 ± 1.42	0.59 ± 0.21	3.96 ± 0.46	0.62 ± 0.09
NN+DN-80 (proposed)	2.43 ± 1.08	0.79 ± 0.12	3.96 ± 0.32	0.71 ± 0.03
140 kVp only
LR-140	2.88 ± 0.29	0.36 ± 0.30	7.27 ± 0.35	−2.97 ± 0.57
NN-140	2.48 ± 1.18	0.73 ± 0.18	2.20 ± 0.40	0.86 ± 0.06

Mean ± SD L1 error and coefficient of determination (R²), averaged across the five cross-validation folds (N=25) and across the five models' predictions on the independent test set (N=10). Best value per column in green.

Bar chart of R² for cross-validation, N=25: LR-80 below zero, NN-80 at 0.59, NN+DN-80 at 0.79, LR-140 at 0.36, NN-140 at 0.73. — R² by model, on cross-validation (left) and the external test set (right). Teal bars use 80 kVp input, orange bars 140 kVp. Both linear baselines fall below zero on the test set and are marked "<0".

Bar chart of R² for the test set, N=10: LR-80 below zero, NN-80 at 0.62, NN+DN-80 at 0.71, LR-140 below zero, NN-140 at 0.86. — R² by model, on cross-validation (left) and the external test set (right). Teal bars use 80 kVp input, orange bars 140 kVp. Both linear baselines fall below zero on the test set and are marked "<0".

The denoiser is what makes ultra-low dose work. Adding it to the 80 kVp pipeline lifts R² from 0.59 to 0.79 in cross-validation and from 0.62 to 0.71 on the test set — a consistent improvement in both validation and testing. Qualitatively (see the figure at the top of this page), NN+DN-80 substantially improves spatial consistency and reduces artifacts relative to NN-80, producing maps that closely match the ground truth and performing comparably to the 140 kVp model while relying on 80 kVp input alone.

The 140 kVp model (NN-140) remains the strongest on the external test set at R² = 0.86, which is expected — it is given a less noisy scan at roughly double the dose. The proposed model's slightly lower test performance may reflect differences in data distribution between the training/validation and test sets, along with the small size of the external cohort.

The point is not that 80 kVp beats 140 kVp — it is that 80 kVp gets close enough, at half the dose of an already ultra-low dose scan, and below that of a two-view chest X-ray.

Discussion & outlook

We presented a novel neural framework for non-invasive liver fat fraction quantification from ultra-low dose CT. Neural fat fraction regression consistently outperforms the standard linear fitting that is the current clinical method, and adding a neural denoiser improves accuracy further — enabling reliable prediction at an effective dose of merely 0.075–0.15 mSv, equivalent to an AP/LAT chest X-ray. To the best of our knowledge, this is the first demonstration of CT-based liver fat fraction assessment at such a low radiation level.

The main limitation is the size of the dataset. With just 35 subjects, performance estimates still carry considerable uncertainty. A large-scale validation study with a substantially broader cohort is under way, though recruitment is slow because the prospective protocol requires both ultra-low dose CT and MRI. Increasing data diversity, especially with respect to liver disease, is expected to yield more robust models. Beyond that we will investigate alternative denoising and 3D regression backbones, and perform systematic hyper-parameter tuning and ensembling.

By showing that a neural regression network can derive accurate liver fat fraction from ultra-low dose CT, this approach paves the way for safer, more accessible population screening for NAFLD. Integrated into radiology workflows, automated fat fraction reporting could support early diagnosis and enable lifestyle interventions or medical treatment before the disease progresses.

Code: lilacharbel/ultra-low-dose-CT-to-3D-liver-fat-fraction

References

Loomba, R., Friedman, S.L. and Shulman, G.I., 2021. Mechanisms and disease consequences of nonalcoholic fatty liver disease. Cell, 184(10), pp.2537–2564.
Nassir, F., Rector, R.S., Hammoud, G.M. and Ibdah, J.A., 2015. Pathogenesis and prevention of hepatic steatosis. Gastroenterology & Hepatology, 11(3), p.167.
Pickhardt, P.J., Graffy, P.M., Reeder, S.B., Hernando, D. and Li, K., 2018. Quantification of liver fat content with unenhanced MDCT: phantom and clinical correlation with MRI proton density fat fraction. American Journal of Roentgenology, 211(3), pp.W151–W157.
Graffy, P.M., Sandfort, V., Summers, R.M. and Pickhardt, P.J., 2019. Automated liver fat quantification at nonenhanced abdominal CT for population-based steatosis assessment. Radiology, 293(2), pp.334–342.
Starekova, J., Hernando, D., Pickhardt, P.J. and Reeder, S.B., 2021. Quantification of liver fat content with CT and MRI: state of the art. Radiology, 301(2), pp.250–262.
Zhang, K., Zuo, W., Chen, Y., Meng, D. and Zhang, L., 2017. Beyond a Gaussian denoiser: residual learning of deep CNN for image denoising. IEEE Transactions on Image Processing, 26(7), pp.3142–3155.
Hatamizadeh, A., Nath, V., Tang, Y., Yang, D., Roth, H.R. and Xu, D., 2021. Swin UNETR: Swin transformers for semantic segmentation of brain tumors in MRI images. In International MICCAI Brainlesion Workshop, pp.272–284.
Johnson, J., Alahi, A. and Fei-Fei, L., 2016. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision – ECCV 2016, pp.694–711.
Green, M., Marom, E.M. and Mayer, A., 2024. Perceptual transformer loss for the neural denoising of ultra-low dose CT. In 2024 IEEE International Symposium on Biomedical Imaging (ISBI), pp.1–5.
Liu, C.K., Liu, C.C., Yang, C.H. and Huang, H.M., 2021. Generation of brain dual-energy CT from single-energy CT using deep learning. Journal of Digital Imaging, 34(1), pp.149–161.

Content on this page is drawn from the paper “MRI-Supervised Ultra-Low Dose CT for 3D Liver Fat Fraction Mapping” (L. Arbel Issaschar, L. Carl, N. Kiryati, M. Green, E. M. Marom, A. Mayer), presented at the MICCAI 2025 Workshop on Machine Learning in Clinical Decision Support (ML-CDS), and the accompanying project presentation.