VCRE-Fib: View-Conditioned Regional Evidence for Fine-Grained Ultrasound Grading of Schistosoma japonicum-Associated Liver Fibrosis

Ziyang Xu; Shuli An; Hao Zhou; Haitian Zhong; Tingting Wu; Tao Wang; Kun Yang; Tieyong Zeng

cs.CVSep 26, 2026

VCRE-Fib: View-Conditioned Regional Evidence for Fine-Grained Ultrasound Grading of Schistosoma japonicum-Associated Liver Fibrosis

Authors: Ziyang Xu, Shuli An, Hao Zhou, Haitian Zhong, Tingting Wu, Tao Wang, Kun Yang, Tieyong Zeng

Organizations: Department of Mathematics, The Chinese University of Hong Kong, Hong Kong, China · Jiangsu Provincial Key Laboratory on Parasite and Vector Control Technology, National Health Commission Key Laboratory of Parasitic Disease Control and Prevention, Jiangsu Institute of Parasitic Diseases, Wuxi, Jiangsu, China · New Laboratory of Pattern Recognition (NLPR), State Key Laboratory of Multimodal Artificial Intelligence Systems (MAIS), Institute of Automation, Chinese Academy of Sciences · Zhongguancun Academy · School of Science, Nanjing University of Posts and Telecommunications, Nanjing, China · Institute for Advanced Study, Beijing Normal-Hong Kong Baptist University, Zhuhai, China · School of Mathematics and Statistics, Guangzhou Nanfang College, Guangzhou, China

Abstract

Accurate assessment of Schistosoma japonicum-associated liver fibrosis is essential for disease management and long-term follow-up in endemic regions. Ultrasound provides non-invasive imaging, but complex local echogenic patterns and anatomical structures make fine-grained grading challenging. Existing deep learning methods can predict fibrosis scores, yet directly incorporating acquisition views and regional cues into grading while retaining spatial information for inspection remains an open problem. Here we present VCRE-Fib, a view-conditioned regional evidence framework that integrates anatomical context, local information, and global image assessment for fine-grained ultrasound grading. The framework forms view-conditioned local grading evidence before spatial pooling, uses weak localization to guide its aggregation, and combines it with global predictions. Image-only inference jointly returns a fibrosis score, acquisition view, and candidate abnormal-region map. We developed and evaluated the method on a re-curated cohort of 108,709 ultrasound images from 6,373 patients across 35 centers. On a patient-disjoint test set of 4,107 images from 240 patients across four centers, VCRE-Fib reduced the prespecified composite grading risk by 7.115% relative to SFibAI trained and evaluated on the same data split. Image-level mean absolute error decreased from 0.391 to 0.378, alongside lower patient-max, patient-median, and center-balanced risks. The full model also achieved lower composite grading risk than variants that separately removed view conditioning or weak localization. These results support incorporating anatomical context and regional evidence into ultrasound grading while exposing spatial predictions for inspection alongside severity estimates.

Figures & tables

Figure 1: VCRE-Fib: view conditioning before spatial aggregation. (a) A shared encoder supplies global context, view probabilities, weak localization and local features. (b) Six view-conditioned paths produce regional logit corrections, weighted by localization and learned support. Conditional distributions share context logits and are mixed by detached view probabilities. (c) One test case and test results from frozen, validation-selected checkpoints. Feature cells are schematic; $\operatorname{sg}$ denotes stop-gradient. Reference boxes are not inference inputs, and localization does not represent exhaustive segmentation.

Method	View module	Weak-localization module	Outputs
SFibAI	Absent	Absent	Score
w/o View	Removed	Retained	Score + region
w/o Weak Loc.	Retained	Removed	Score + view
VCRE-Fib	Retained	Retained	Score + view + region

Table 1: Compared models and inference outputs. All use image-only inference, the same backbone initialization and grading objective. Deletion variants retain the regional grading path and are trained independently.

Rank	Method	$R_{\mathrm{final}}$	Image COR	Max COR	Median COR	Center COR
1	VCRE-Fib	0.186865	0.177908	0.200528	0.154344	0.177453
2	w/o Weak Loc.	0.191308	0.186343	0.202422	0.165844	0.179008
3	w/o View	0.193799	0.181518	0.210002	0.158449	0.185957
4	SFibAI	0.201180	0.185938	0.219742	0.162587	0.194539

Table 2: Test set primary results: 4,107 images, 240 patients, four centers. All checkpoints were frozen after validation selection. All risks are lower-is-better. Bold denotes the best unrounded value per metric; exact ties are all bold.

Metric	SFibAI	w/o View	w/o Weak Loc.	VCRE-Fib
MAE $\downarrow$	0.391313	0.384734	0.392205	0.377857
RMSE $\downarrow$	0.568137	0.564076	0.567780	0.553435
Acc. $\pm 0.3$ (%) $\uparrow$	56.17	57.51	56.05	57.95
Acc. $\pm 0.5$ (%) $\uparrow$	69.61	70.08	69.47	70.15
4-grade acc. (%) $\uparrow$	64.11	66.13	64.13	67.03
Macro-F1 $\uparrow$	0.639702	0.663674	0.643722	0.671466

Table 3: Image-level test metrics; 4,107 images per model. Bold denotes the best unrounded value per metric; exact ties are all bold.

Metric	SFibAI	w/o View	w/o Weak Loc.	VCRE-Fib
View N	N/A	N/A	4107	4107
View acc. (%) $\uparrow$	N/A	N/A	75.70	66.81
View macro-F1 $\uparrow$	N/A	N/A	0.757500	0.665861
View CE $\downarrow$	N/A	N/A	0.690513	0.916542
Weak-target N	N/A	4035	N/A	4035
[email protected] $\uparrow$	N/A	0.430661	N/A	0.360312

Table 4: Spatial test outputs. N/A indicates an absent head. Weak-target N counts eligible images, not rectangles; Dice/IoU measure agreement with local annotations, not complete lesion segmentation. Bold denotes the best unrounded value per metric; exact ties are all bold.

Figure 2: Four selected test cases, with reference and predicted scores and acquisition views at left. Predictions use the frozen checkpoint evaluated quantitatively. Rows (a)–(c) show localization correspondence; (d) shows a response beyond local boxes. Green boxes denote clinician annotations; localization uses a $[0,1]$ scale, opacity 0.55, and a cyan 0.5 contour. In row (a), the predicted view differs from the reference. These selected examples illustrate spatial outputs without establishing population localization performance or complete lesion extent.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

	Total	Train	Validation	Test
Images	108,709	83,722	20,880	4,107
Patients	6,373	4,906	1,227	240
Centers	35	35	33	4

Appendix

Table S1: Composition of the curated cohort used in this study. Centers overlap between splits; patients do not.

Method	Rank	$R_{\mathrm{final}}\downarrow$	$\Delta$ absolute $\downarrow$	$\Delta$ relative (%) $\downarrow$
VCRE-Fib	1	0.186865	-0.014315	-7.115
w/o Weak Loc.	2	0.191308	-0.009872	-4.907
w/o View	3	0.193799	-0.007380	-3.668
SFibAI	4	0.201180	0.000000	0.000

Appendix

Table S2: Observed test differences relative to SFibAI. Differences are method minus baseline; relative differences divide by the baseline, using unrounded values. Bold denotes the best unrounded value per metric; exact ties are all bold.

Method	Level	MAE $\downarrow$	Acc. $\pm.3\uparrow$	Acc. $\pm.5\uparrow$	4-grade $\uparrow$	TMAE $\downarrow$	Severe $\downarrow$	COR $\downarrow$
SFibAI	Max	0.472421	50.42	63.75	61.25	0.182625	16.25	0.219742
w/o View	Max	0.445159	53.75	66.25	63.75	0.171660	15.42	0.210002
w/o Weak Loc.	Max	0.427688	53.33	67.50	63.75	0.149615	12.50	0.202422
VCRE-Fib	Max	0.420558	53.75	65.42	65.42	0.143967	11.67	0.200528
SFibAI	Median	0.348790	62.08	72.08	67.08	0.102808	7.92	0.162587
w/o View	Median	0.346292	61.25	73.75	72.08	0.108306	7.92	0.158449

Appendix

Table S3: Patient-level test metrics, 240 patients per row. Reference and prediction are aggregated separately; accuracies and severe errors are percentages. Bold compares models within each aggregation.

Metric	SFibAI	w/o View	w/o Weak Loc.	VCRE-Fib
Patient Max RMSE $\downarrow$	0.665765	0.642874	0.611087	0.596914
Patient Max macro-F1 $\uparrow$	0.606287	0.635779	0.636286	0.655596
Patient Median RMSE $\downarrow$	0.508183	0.519102	0.528079	0.505147
Patient Median macro-F1 $\uparrow$	0.675454	0.726897	0.702926	0.721044
Image macro-AUROC $\uparrow$	0.879791	0.886155	0.879378	0.891698
Image macro-AUPRC $\uparrow$	0.699348	0.719365	0.699530	0.729264

Appendix

Table S4: Additional test grading and image-only probability metrics. Bold denotes the best unrounded value per metric; exact ties are all bold.

Center	Images	Patients	SFibAI	w/o View	w/o Weak Loc.	VCRE-Fib
center_05	25	2	0.005361	0.006614	0.005765	0.005329
center_07	1948	115	0.201563	0.192100	0.192990	0.183879
center_15	1599	93	0.308194	0.295332	0.274507	0.281672
center_17	535	30	0.029521	0.027661	0.028223	0.025816

Appendix

Table S5: Per-center test patient-max COR (lower is better). Bold compares methods within a center; small-center estimates remain descriptive.

Method	Batch	Parameters	GMACs/image	Median ms	P95 ms	Peak MiB
SFibAI	1	23581796	21.353	5.620	5.626	154.318
w/o View	1	25311562	22.103	6.171	6.188	160.919
w/o Weak Loc.	1	24832975	21.654	7.122	7.165	159.093
VCRE-Fib	1	25422928	22.258	7.371	7.397	161.344
SFibAI	8	23581796	21.353	30.088	30.119	539.318
w/o View	8	25311562	22.103	31.067	31.133	545.919

Appendix

Table S6: Matched resource measurements; latency is milliseconds per batch and memory is peak allocated MiB. All quantities are lower-is-better; bold compares models within each batch size.

View	Anatomical/acquisition view	N	w/o Weak Loc.	VCRE-Fib
1	Subxiphoid sagittal plane (left lobe of the liver)	694	93.66	84.73
2	Left subcostal transverse plane (left lobe and sagittal segment of portal vein)	699	74.82	65.09
3	Right subcostal plane (second porta hepatis and confluence of three hepatic veins)	689	59.51	54.57
4	Right subcostal plane (right hepatic lobe, right hepatic vein, diaphragm)	671	58.72	52.61
5	Right intercostal oblique plane (right hepatic lobe, portal vein, gallbladder)	682	81.52	66.86
6	Right intercostal oblique plane (right hepatic lobe and right kidney)	672	85.71	76.79

Appendix

Table S7: Six acquisition views and test recall (percent). Bold compares the two view-capable models within each class. Main-paper cases cover four true views.

Figure S1: Selected validation examples with responses beyond local boxes. Predictions use the frozen checkpoint evaluated quantitatively. The three columns show the original ROI, clinician annotation, and model localization; labels report reference and predicted scores and acquisition views. In the first example, the predicted view differs from the reference. Responses beyond local boxes identify locations for inspection, without establishing complete lesion extent.

Method	Epoch cap	Selected epoch	Val $R_{\mathrm{final}}\downarrow$	Val image COR $\downarrow$
SFibAI	90	65	0.172661	0.142376
w/o View	90	80	0.157844	0.134258
w/o Weak Loc.	90	39	0.159315	0.134234
VCRE-Fib	90	47	0.156067	0.134735

Appendix

Table S8: Validation-selected frozen checkpoints. Validation: 20,880 images, 1,227 patients, 33 centers. All four methods, including VCRE-Fib, use a 90-epoch cap for the training comparison; checkpoints are selected from epochs 21–90. Bold denotes the best unrounded value per metric; exact ties are all bold.

Figure S2: Training and validation trajectories under a common 90-epoch cap. (a) Total training loss on a logarithmic axis, including each model’s auxiliary terms. (b) Validation

R_{\mathrm{final}}

. (c) Validation image COR. Curves show unsmoothed epoch-level observations; dots mark the validation-selected checkpoints in Table S8 .

Figure S3: Full VCRE-Fib loss components under the same 90-epoch cap, shown as training diagnostics rather than estimates of module effects.

Explore similar work

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
How Far Has AI Come in Liver Fibrosis Staging? A Large-Scale Real-World Dataset and Benchmark
May 25, 2026Yuanye Liu, Nannan Shi, Zhejia Zhang +20Medical Image Classification Medical Image Benchmarks
School of Data Science, Fudan University, Shanghai, China · Department of Radiology, Shanghai Public Health Clinical Center, Fudan University, Shanghai, China · Department of Electrical and Computer Engineering, Northwestern University, Evanston, USA +12
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

Sep 3, 2026J/K move · Enter open · S save

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

Guangyi Zhang, Xiaohong Wang, Eugene Cheah, Peng Guo+3

Center for Ultrasound Research & Translation, Massachusetts General Hospital, Harvard Medical School, Boston, MA, USA · Lincoln Laboratory, Massachusetts Institute of Technology, Lexington, MA, USA

Ultrasound Imaging Medical Image Classification Clinical Risk Prediction

Open paper AskPDF arXiv

Metabolic dysfunction-associated steatotic liver disease (MASLD) affects approximately 30% of the general population. Ultrasound-based imaging, including B-mode imaging and shear wave elastography (SWE), is widely used for noninvasive fibrosis assessment; however, the role of deep learning-based ultrasound image learning for MASLD risk stratification remains insufficiently characterized. In this study, we developed and evaluated ultrasound image learning pipelines using B-mode and SWE images for fibrosis staging and identification of patients with at-risk metabolic dysfunction-associated steatohepatitis (MASH). A total of 250 ultrasound examinations, one exam per subject, were included. Model performance was evaluated using 3-fold cross-validation with area under the receiver operating characteristic curve (AUROC). End-to-end SWE image learning achieved performance comparable to operator-guided SWE across fibrosis stages. Overall, SWE-based learning consistently outperformed B-mode image learning in fibrosis staging, with AUROC improvements from 0.64 (95%CI: [0.56, 0.72]) to 0.72 (95% CI: [0.65, 0.79]) for F>=2 (significant fibrosis, p=0.11), from 0.67 (95%CI: [0.58, 0.75]) to 0.78 (95% CI:[0.72, 0.85]) for F>=3 (advanced fibrosis, p=0.02), and from 0.69 (95%CI: [0.56, 0.82]) to 0.80 (95%CI: [0.72, 0.89]) for F4 (cirrhosis, p=0.10). These findings highlight the potential of SWE image learning for MASLD risk stratification.

Abstract

Explore similar work

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

How Far Has AI Come in Liver Fibrosis Staging? A Large-Scale Real-World Dataset and Benchmark

Learning from Complementary Ultrasound Representations for Liver Disease Classification