UNICORN: Ultrasound Nakagami Imaging via Score Matching and Adaptation for Assessing Hepatic Steatosis

Kwanyoung Kim; Jaa-Yeon Lee; Youngjun Ko; GunWoo Lee; Jong Chul Ye

eess.IVMar 16, 2026

UNICORN: Ultrasound Nakagami Imaging via Score Matching and Adaptation for Assessing Hepatic Steatosis

Authors: Kwanyoung Kim, Jaa-Yeon Lee, Youngjun Ko, GunWoo Lee, Jong Chul Ye

Organizations: Department of AI Convergence, GIST · Kim Jae Chul Graduate School of AI, KAIST · Samsung Medison

Abstract

Ultrasound imaging is an essential first-line tool for assessing hepatic steatosis. While conventional B-mode ultrasound imaging has limitations in providing detailed tissue characterization, ultrasound Nakagami imaging holds promise for visualizing and quantifying tissue scattering in backscattered signals, with potential applications in fat fraction analysis. However, existing methods for Nakagami imaging struggle with optimal window size selection and suffer from estimator instability, leading to degraded image resolution. To address these challenges, we propose a novel method called UNICORN (Ultrasound Nakagami Imaging via Score Matching and Adaptation), which offers an accurate, closed-form estimator for Nakagami parameter estimation based on the score function of the ultrasound envelope signal. Unlike methods that visualize only specific regions of interest (ROI) and estimate parameters within fixed window sizes, our approach provides comprehensive parameter mapping by providing a pixel-by-pixel estimator, resulting in high-resolution imaging. We demonstrated that our proposed estimator effectively assesses hepatic steatosis and provides visual distinction in the backscattered statistics associated with this condition. Through extensive experiments using real envelope data from patient, we validated that UNICORN enables clinical detection of hepatic steatosis and exhibits robustness and generalizability.

Figures & tables

Fig. 1 : Nakagami imaging using conventional methods and our UNICORN framework. (a) Momentum-based approach uses a sliding window for Nakagami parameter calculation. (b) Maximum likelihood method obtains Nakagami image through ML estimation. (c) UNICORN consists of two stages: training a score model to learn RF envelope score function, and inference step estimates Nakagami image in terms of score function.

	MRI-PDFF
Dataset	$<5\%$ (Normal)	$5\%$ - $15\%$ (Mild)	$\geq 15\%$ (Severe)	Total
SMC	224	170	55	449

TABLE I : The number of subjects across hepatic steatosis stages determined by MRI PDFF on each dataset. One frame per patient is used for statistical analysis.

Fig. 2 : Comparison of qualitative results across various methods: (Row 1) MNIST dataset, (Rows 2-3) BUSI Ultrasound Image dataset. Comparison against Momentum, MLE, WMC, and UNICORN. The yellow numbers indicate PSNR. To visualize the results, we normalize the images and convert them to grayscale.

Dataset	MNIST			BUSI
Metric	WS	PSNR (dB) $\uparrow$	RMSE $\downarrow$	WS	PSNR (dB) $\uparrow$	RMSE $\downarrow$
Measurement	-	9.17	0.695	-	10.14	0.651
Momentum	9	21.01	0.179	9	18.58	0.230
Momentum	11	21.28	0.173	13	21.22	0.174
Momentum	13	20.78	0.184	25	23.31	0.132
MLE	11	20.45	0.191	25	22.65	0.148

TABLE II : Quantitative simulation results for MNIST dataset and ultrasound image BUSI dataset using various methods. The bold numbers indicate the best performance. WS denotes sliding window size.

Fig. 3 : Comparison of qualitative results obtained using different methods. Each row corresponds to Normal, Mild, and Severe liver cases, with MRI-PDFF values of 0.71, 6.34, and 22.12, respectively. The comparison includes the Momentum, MLE, WMC, and UNICORN methods. Red dashed lines delineate the liver region. Compared to other methods, our approach provides clearer visual cues for distinguishing between normal, mild, and severe fatty liver cases.

Fig. 4 : Scatter plots illustrating the relationship between MRI-PDFF and estimated m parameters across various methods and datasets: Columns 1-4 represent Momentum, MLE, WMC, and UNICORN, respectively; PCC denotes the Pearson Correlation Coefficient between MRI-PDFF and the estimated $m$ parameter. The trend line follows a locally weighted linear regression model.

Method	SMC
Method	PCC(r)	SS
Momentum	0.53
MLE	0.65
WMC	0.54
UNICORN(Ours)	0.77

TABLE III : Comparison of Pearson Correlation Coefficient (PCC) between MRI-PDFF and estimated Nakagami

m

parameters and Statistical Significance (SS) across various methods

Fig. 5 : Box plots illustrating the relationship between MRI-PDFF and estimated m parameters across different methods at each stage: Methods include Momentum, MLE, WMC, and UNICORN; stages are defined as Normal (MRI-PDFF $<$ 5 $\%$ ), Mild (5 $\%$ $\leq$ MRI-PDFF $<$ 15 $\%$ ), and Severe (MRI-PDFF $>$ 15 $\%$ ). The $p$ -values between neighboring stages are indicated within the box plots.

Fig. 6 : ROC curves for estimated m parameters in the differentiation of hepatic steatosis stages using overall datasets across different methods: Momentum, MLE, WMC, and UNICORN. Stages are defined as Normal (MRI-PDFF $<$ 5 $\%$ ), Mild (5 $\%$ $\leq$ MRI-PDFF $<$ 15 $\%$ ), and Severe (MRI-PDFF $>$ 15 $\%$ ). (A) Normal vs Mild, (B) Mild vs Severe, and (c) Normal vs Severe.

	AUC	Sensitivity	Specificity	PPV	NPV	F1
(A) Normal vs Mild	0.87
(a) Optimal		0.73	0.87	0.81	0.80	0.77
(b) 90% Specificity		0.68	0.90	0.85	0.78	0.76
(c) 90% Sensitivity		0.90	0.57	0.62	0.88	0.74
(B) Mild vs Severe	0.80
(a) Optimal		0.83	0.62	0.42	0.92	0.55

TABLE IV : ROC curve analysis of UNICORN ( Fig. 6 ). (a) optimal threshold, (b) threshold for 90% specificity, (b) threshold for 90% sensitivity. PPV denotes positive predictive value and NPV denotes negative predictive value.

	Method	AUC	Sensitivity	Specificity	PPV	NPV	F1
(A) Normal vs Mild	Momentum	0.83	0.85	0.68	0.68	0.85	0.76
	MLE	0.88	0.83	0.78	0.75	0.85	0.79
	WMC	0.82	0.88	0.65	0.68	0.87	0.76
	UNICORN	0.87	0.73	0.87	0.81	0.80	0.77
(B) Mild vs Severe	Momentum	0.68	0.70	0.60	0.36	0.86	0.47
	MLE	0.74	0.68	0.67	0.41	0.87	0.51

TABLE V : ROC curve analysis comparing baseline methods corresponding to Fig. 6 . The threshold is selected by the highest F1 score. PPV denotes positive predictive value, and NPV denotes negative predictive value.

Fig. 7 : Comparison of with and without low-pass filter. (a) represents normal liver cases with MRI-PDFF values of 4.2, and (b) represents fatty liver cases with MRI-PDFF values of 30.4. The red dashed lines indicate the liver region.

Method	Momentum	MLE	WMC	UNICORN
Train Memory (GB)	-	-	-	13.68
Duration (sec/patient)	0.084 $\pm$ 0.003	33.79 $\pm$ 2.451	0.131 $\pm$ 0.005	0.004 $\pm$ 0.001

TABLE VI : Computational complexity including training memory latency and inference speed.

Ultrasound Imaging

Ultrasound imaging uses sound waves to visualize anatomy and function, with AI improving image reconstruction, diagnosis, measurement, and scan guidance.

See all 150 papers

Learning from Complementary Ultrasound Representations for Liver Disease Classification
Jul 13, 2026Sabahattin Mert Daloglu, Gokce Bekar, Ceren Coskun +5Medical Image Classification Ultrasound Imaging
PONS Incorporated, Newark, NJ, USA · Department of Radiology, Mayo Clinic, Jacksonville, FL, USA · Department of Radiology, Department of Medicine, University of British Columbia, Vancouver, Canada
Development and Evaluation of Ultrasound Image Learning Pipelines for MASLD Risk Stratification
Sep 3, 2026Guangyi Zhang, Xiaohong Wang, Eugene Cheah +4Ultrasound Imaging Medical Image Classification
Center for Ultrasound Research & Translation, Massachusetts General Hospital, Harvard Medical School, Boston, MA, USA · Lincoln Laboratory, Massachusetts Institute of Technology, Lexington, MA, USA
A Unified Framework for the Detection and Classification of Fatty Pancreas in Ultrasound Images
May 8, 2026Ioan-Tudor-Alexandru Anghel, Ciprian-Mihai Ceausescu, Elena Dana Nedelcu +5Medical Image Classification Ultrasound Image Classification
Faculty of Mathematics and Computer Science, University of Bucharest, Bucharest, Romania · Ponderas Academic Hospital, Radiology, Bucharest, Romania · Ponderas Academic Hospital, Internal Medicine, Bucharest, Romania +2

Jul 13, 2026J/K move · Enter open · S save

Learning from Complementary Ultrasound Representations for Liver Disease Classification

Sabahattin Mert Daloglu, Gokce Bekar, Ceren Coskun, Senanur Sahin+4

PONS Incorporated, Newark, NJ, USA · Department of Radiology, Mayo Clinic, Jacksonville, FL, USA · Department of Radiology, Department of Medicine, University of British Columbia, Vancouver, Canada

Medical Image Classification Ultrasound Imaging Medical Image Representation Learning Ultrasound Image Classification

Open paper AskPDF arXiv

Differentiating non-alcoholic steatohepatitis (NASH) from non-alcoholic fatty liver disease (NAFLD) using ultrasound remains challenging due to subtle tissue alterations and the limited information available in conventional B-mode imaging. In this work, we investigate whether complementary ultrasound representations derived from the same acquisition can improve NASH versus NAFLD classification. Specifically, we combine conventional B-mode ultrasound with physics-guided and local phase-based image representations and evaluate their effectiveness using self-supervised masked autoencoders (MAEs) and graph convolutional networks (GCNs). Experiments were conducted on a multi-site Mayo Clinic cohort consisting of 2,547 liver ultrasound scans from 125 patients. Compared with conventional B-mode ultrasound alone, complementary ultrasound representations consistently improved classification performance, yielding gains of up to 32.4% in accuracy and 91.2% in F1-score. Furthermore, performance improvements were consistently observed across age groups, sex, race, ethnicity,and acquisition sites.

Abstract

Ultrasound Imaging

Explore similar work

Learning from Complementary Ultrasound Representations for Liver Disease Classification

Development and Evaluation of Ultrasound Image Learning Pipelines for MASLD Risk Stratification

A Unified Framework for the Detection and Classification of Fatty Pancreas in Ultrasound Images