UniField++: A Unified Agentic Framework with Loop Reasoning for MRI Field-Strength Enhancement

Yiyang Lin; Chenhui Wang; Tianliang Yao; Pengli Zhu; Shuaibo Li; Zijun Sun; Yiming Deng; Yixuan Yuan

cs.CVOct 16, 2026

UniField++: A Unified Agentic Framework with Loop Reasoning for MRI Field-Strength Enhancement

Authors: Yiyang Lin, Chenhui Wang, Tianliang Yao, Pengli Zhu, Shuaibo Li, Zijun Sun, Yiming Deng, Yixuan Yuan

Organizations: Department of Electronic Engineering, The Chinese University of Hong Kong, Hong Kong SAR, China · Institute of Science and Technology for Brain-inspired Intelligence, Fudan University, Shanghai 200433, China · Institute of Diagnostic Radiology, University Hospital Zurich, University of Zurich, Zurich 8091, Switzerland · Department of Neurology, Beijing Tiantan Hospital, Capital Medical University, Beijing 100070, China ·

Abstract

MRI field-strength enhancement seeks to combine the accessibility of low-field scanners with the image quality of high-field scanners. Yet, existing methods remain prone to anatomical hallucinations, as uniform processing leaves difficult brain regions insufficiently examined and scarce paired data limit their learning of how tissue contrast and anatomical visibility change across field strengths. To address these challenges, we propose UniField++, a two-agent framework that unifies enhancement tasks across multiple field-strength transitions, including 64mT-to-3T and 3T-to-7T. Unified learning allows both the VLM-based planner and the flow-matching-based enhancer to draw on field-dependent image characteristics observed across tasks, rather than relying on a single transition. Trained with radiologists' annotations of hallucination-prone regions and corresponding correction plans, the planner predicts regional risk and formulates targeted restoration strategies. Guided by these plans, the enhancer performs loop reasoning along the flow-matching trajectory, repeatedly processing high-risk regions for additional refinement. Together, cross-transition learning and risk-guided refinement support anatomically faithful enhancement across field strengths. Across the evaluated tasks, UniField++ improves PSNR by an average of 2.74 dB and SSIM by 14.05 percentage points over comparison methods, with corresponding gains of 3.97 dB and 13.92 percentage points over the ablation baseline. It also maintains enhancement quality on an external cohort with an unseen field strength and improves downstream whole-brain segmentation Macro Dice by an average of 0.1457 over the original low-field images. Code is available at https://github.com/CUHK-AIM-Group/UniField_plusplus.

Figures & tables

Fig. 1: Overview of UniField++. Our framework unifies enhancement tasks across various types of field strengths. And UniField++ primarily comprises a VLM-based planner that simulates the radiologists’ planning process for enhancement, and a flow-matching-based enhancer that executes the enhancement according to the plan.

Fig. 2: Schematic diagram of UniField++. Our framework mainly consists of two parts: (a) In the construction of Radiologist-Informed Enhancement Planner , (1) for the Hallucination Annotating, radiologists identify hallucinations in MRI field strength enhancement and provide restoration plans; (2) for the Hallucination Learning, a Vision-Language Model (VLM) based planner is fine-tuned to predict hallucinations and generate the restoration plans; (b) In the Unified Agentic Enhancement with Loop Reasoning , guided by the plans provided by the VLM, a flow-matching based enhancer containing the loop transformer is utilized for MRI field strength enhancement.

Task	Dataset	Modalities	Registered	Total	Train	Test
64mT-to-3T	ULF-EnC [ 5 ]	T1, T2, FLAIR	Yes	50	40	10
	Leiden [ 37 ]	T1, T2, FLAIR	No	10	8	2
	KCL [ 7 ]	T1, T2	No	23	18	5
3T-to-7T	UNC [ 8 ]	T1, T2	Yes	10	8	2
3T-to-7T	BNU [ 9 ]	T1, T2	No	20	16	4
50mT-to-3T	BTH	T1, T2, FLAIR	Yes	30	0	30

TABLE I: Dataset overview. Three public datasets for 64mT-to-3T, two for 3T-to-7T, and a private 50mT-to-3T dataset are used.

Fig. 3: Qualitative comparison on the internal dataset for 64mT-to-3T and 3T-to-7T across multiple modalities. The value in the lower-right corner of each image is the mean absolute error (MAE).

Methods	T1				T2				FLAIR
Methods	NRMSE $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	NRMSE $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	NRMSE $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
MO-U-NET (HBM’25)	11.4	18.98	52.9	0.278	9.5	20.58	59.2	0.233	10.9	19.32	53.3	0.249
MSFA (MICCAI’24)	11.3	19.05	67.9	0.255	9.9	20.27	71.6	0.220	10.9	19.31	65.6	0.238
LowGAN (Radiology’25)	12.8	18.03	66.6	0.195	11.4	18.98	65.8	0.195	14.1	17.02	63.2	0.195
FlashVSR (CVPR’26)	12.1	18.41	63.4	0.298	12.0	18.54	66.3	0.235	15.1	16.61	60.6	0.300
FCVSR (TMM’26)	14.1	17.16	61.7	0.258	11.9	18.63	64.9	0.217	16.6	15.67	55.6	0.258

TABLE II: 64mT-to-3T comparison for T1, T2, and FLAIR. NRMSE and SSIM are in %, PSNR in dB, and LPIPS is unitless. Bold and underlined values indicate the best and second-best results.

Methods	MO-U-NET HBM ’25	MSFA MICCAI ’24	LowGAN Radiology ’25	FlashVSR CVPR ’26	FCVSR TMM ’26	UniField MICCAI ’26	UniField++ (Ours)
T1
NRMSE $\downarrow$	11.9	10.5	15.4	11.8	14.2	9.9	8.2
PSNR $\uparrow$	18.51	19.66	16.26	18.63	16.97	20.12	21.79
SSIM $\uparrow$	60.8	66.8	54.0	63.0	56.8	66.1	77.1
LPIPS $\downarrow$	0.277	0.234	0.248	0.228	0.264	0.207	0.198
T2

TABLE III: 3T-to-7T comparison for T1 and T2. Bold and underlined values indicate the best and second-best results.

Fig. 4: Qualitative comparison on the external dataset for the 50mT-to-3T task.

Methods	T1				T2				FLAIR
Methods	NRMSE $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	NRMSE $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	NRMSE $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
MO-U-NET (HBM’25)	9.7	20.32	76.6	0.212	9.0	21.11	79.6	0.184	12.5	18.10	74.2	0.227
MSFA (MICCAI’24)	10.1	20.05	77.2	0.174	9.2	20.87	79.3	0.174	13.9	17.21	74.2	0.215
LowGAN (Radiology’25)	10.0	20.12	73.6	0.174	11.5	18.84	72.6	0.180	13.7	17.28	70.8	0.191
FlashVSR (CVPR’26)	9.2	20.88	75.7	0.249	8.8	21.32	78.1	0.176	12.2	18.34	74.6	0.232
FCVSR (TMM’26)	10.5	19.75	75.7	0.190	9.3	20.77	78.8	0.158	11.7	18.69	75.1	0.179

TABLE IV: External quantitative results on the 50mT-to-3T task. Bold and underlined values denote the best and second-best results, respectively; underlining is omitted when the best result is tied.

Fig. 5: Qualitative ablation of components in UAELR. Column labels: P, plan; A, atlas; L, loop; F, decoder fine-tuning. Red boxes and zoom-ins highlight the changes introduced by each component.

Setting	Region F1 $\uparrow$	Issue F1 $\uparrow$	Plan F1 $\uparrow$	Semantic $\uparrow$
W/o HL	0.6546	0.3986	0.6833	0.3548
With HL	0.8511	0.6383	0.8313	0.5213

TABLE V: Hallucination learning (HL) ablation.

Planner		Enhancer		Hyper- parameters	Setting	T1				T2				FLAIR
Plan	Atlas	Loop	Finetune decoder	Hyper- parameters	Setting	NRMSE $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	NRMSE $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	NRMSE $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
$\times$	$\times$	$\times$	$\times$	–	–	13.1	17.77	68.1	0.252	14.7	16.82	60.9	0.268	14.4	16.99	63.5	0.282
$\checkmark$	$\times$	$\times$	$\times$	–	–	10.9	19.38	71.2	0.194	8.7	21.26	71.6	0.171	10.0	20.05	67.0	0.212
$\checkmark$	$\checkmark$	$\times$	$\times$	\Block 3-1 $\lambda_{\mathrm{risk}}$	1	10.9	19.40	72.9	0.190	9.0	21.03	74.1	0.163	10.1	19.95	69.4	0.206
					5	10.7	19.47	72.7	0.184	8.8	21.21	74.1	0.164	9.6	20.40	70.9	0.197
					10	10.9	19.31	70.0	0.193	8.7	21.25	73.1	0.171	10.8	19.45	66.9	0.210

TABLE VI: Ablation results on the 64mT-to-3T task. NRMSE and SSIM are in %, PSNR in dB, and LPIPS is unitless. Bold and underlined values indicate the best and second-best results, respectively; underlining is omitted when the best result is tied.

\lambda_{\mathrm{risk}}

is the risk weight in Eq. ( 9 ); Loop block number specifies the blocks used for cyclic inference in the 30-block backbone.

Planner		Enhancer		Hyper- parameters	Setting	T1				T2
Plan	Atlas	Loop	Finetune decoder	Hyper- parameters	Setting	NRMSE $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	NRMSE $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
$\times$	$\times$	$\times$	$\times$	–	–	12.4	18.22	60.7	0.261	10.5	19.66	62.7	0.231
$\checkmark$	$\times$	$\times$	$\times$	–	–	11.3	18.96	64.4	0.210	8.4	21.61	65.3	0.188
$\checkmark$	$\checkmark$	$\times$	$\times$	\Block 3-1 $\lambda_{\mathrm{risk}}$	1	10.0	20.05	67.9	0.206	8.3	21.69	68.5	0.178
					5	10.2	19.92	67.0	0.198	7.5	22.48	71.9	0.173
					10	9.6	20.49	68.7	0.200	8.1	21.87	70.7	0.178

TABLE VII: Ablation results on the 3T-to-7T task. Bold and underlined values indicate the best and second-best results, respectively; underlining is omitted when the best result is tied.

\lambda_{\mathrm{risk}}

is the risk weight in Eq. ( 9 ); Loop block number specifies the blocks used for cyclic inference in the 30-block backbone.

Task	Modality	$n_{\mathrm{test}}$	Macro Dice $\uparrow$	Foreground Dice $\uparrow$	Macro IoU $\uparrow$	Volume Similarity $\uparrow$
Task	Modality	$n_{\mathrm{test}}$	(Input / Virtual)	(Input / Virtual)	(Input / Virtual)	(Input / Virtual)
\Block 3-164mT $\rightarrow$ 3T	T1	17	0.5366 / 0.7803	0.8573 / 0.9205	0.3921 / 0.6538	0.9315 / 0.9723
	T2	17	0.7220 / 0.8350	0.9082 / 0.9404	0.5788 / 0.7226	0.9617 / 0.9874
	FLAIR	12	0.4396 / 0.7793	0.8343 / 0.9317	0.3127 / 0.6486	0.8925 / 0.9828
\Block 2-13T $\rightarrow$ 7T	T1	6	0.8294 / 0.8412	0.9420 / 0.9468	0.7169 / 0.7333	0.9852 / 0.9898
	T2	6	0.7747 / 0.7950	0.9302 / 0.9359	0.6446 / 0.6710	0.9723 / 0.9752

TABLE VIII: Downstream results of UniField++. Metrics are shown for input / virtual high-field MRI, using SynthSeg on real high-field MRI as the reference.

n_{\mathrm{test}}

is the number of test volumes per modality.

Fig. 6: Qualitative downstream segmentation results. The numbers below the segmentation maps are Dice scores.

Magnetic Resonance Imaging

Magnetic resonance imaging uses magnetic fields and radio waves to create images of anatomy and tissue function for research and clinical care.

See all 251 papers

UniField: A Unified Field-Aware MRI Enhancement Framework
Mar 10, 2026Yiyang Lin, Chenhui Wang, Zhihao Peng +1Medical Image Super-Resolution Magnetic Resonance Imaging
Department of Electronic Engineering, the Chinese University of Hong Kong · The Institute of Science and Technology for Brain-inspired Intelligence, Fudan University
Enhancing Ultra-low-field MRI with Segmentation-guided Adversarial Learning
May 27, 2026James Grover, Andrew Phair, Michael Ferraro +1Image Enhancement Magnetic Resonance Imaging
Image X Institute, Sydney School of Health Sciences, Faculty of Medicine and Health, The University of Sydney
MRI Super-Resolution with RCDM/WaveMix and Task-Aware Segmentation
Sep 30, 2026Kavitha Viswanathan, Harsh Choudhary, Amit SethiMedical Image Analysis Medical Image Super-Resolution
Department of Electrical Engineering, Indian Institute of Technology Bombay, Mumbai 400 076, India

Mar 10, 2026J/K move · Enter open · S save

UniField: A Unified Field-Aware MRI Enhancement Framework

Yiyang Lin, Chenhui Wang, Zhihao Peng, Yixuan Yuan

Department of Electronic Engineering, the Chinese University of Hong Kong · The Institute of Science and Technology for Brain-inspired Intelligence, Fudan University

Medical Image Super-Resolution Magnetic Resonance Imaging

Open paper AskPDF arXiv

Magnetic Resonance Imaging (MRI) field-strength enhancement holds immense value for both clinical diagnostics and advanced research. However, existing methods typically focus on isolated enhancement tasks, such as specific 64mT-to-3T or 3T-to-7T transitions using limited subject cohorts, thereby failing to exploit the shared degradation patterns inherent across different field strengths and severely restricting model generalization. To address this challenge, we propose \methodname, a unified framework integrating multiple modalities and enhancement tasks to mutually promote representation learning by exploiting these shared degradation characteristics. Specifically, our main contributions are threefold. Firstly, to overcome MRI data scarcity and capture continuous anatomical structures, \methodname departs from conventional methods that treat 3D MRI volumes as independent 2D slices. Instead, we directly exploit comprehensive 3D volumetric information by leveraging pre-trained 3D foundation models, thereby embedding generalized and robust structural representations to significantly boost enhancement performance. In addition, to mitigate the spectral bias of mainstream flow-matching models that often over-smooth high-frequency details, we explicitly incorporate the physical mechanisms of magnetic fields to introduce a Field-Aware Spectral Rectification Mechanism (FASRM), tailoring customized spectral corrections to distinct field strengths. Finally, to resolve the fundamental data bottleneck, we organize and publicly release a comprehensive paired multi-field MRI dataset, which is an order of magnitude larger than existing datasets. Extensive experiments demonstrate our method's superiority over state-of-the-art approaches, achieving an average improvement of approximately 1.81 dB in PSNR and 9.47% in SSIM. Codes and datasets are available at: https://github.com/linyiyang98/UniField.

Abstract

Magnetic Resonance Imaging

Explore similar work

UniField: A Unified Field-Aware MRI Enhancement Framework

Enhancing Ultra-low-field MRI with Segmentation-guided Adversarial Learning

MRI Super-Resolution with RCDM/WaveMix and Task-Aware Segmentation