Triple-Phase Sequential Fusion Network with Contrast Enhancement Kinetics for Hepatobiliary-Phase Liver MRI Synthesis

arXiv:2604.22904 · eess.IV, cs.CV · Submitted 2026-04-24 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: We're looking at the paper: "Triple-Phase Sequential Fusion Network with Contrast Enhancement Kinetics for Hepatobiliary-Phase Liver MRI Synthesis".

Dev: Here's the gist of the paper. Triple-Phase Sequential Fusion Network (TriPF-Net) is proposed to synthesize hepatobiliary phase (HBP) images by leveraging sequential information from pre-HBP sequences to reduce reliance on prolonged post-contrast delays The Problem Addressed Gadoxetate disodium-enhanced MRI is essential for the detection and characterization of hepatocellular carcinoma, but acquiring the hepatobiliary phase (HBP) requires a prolonged post-contrast delay which reduces workflow efficiency and increases the risk of motion artifacts.

Rosa: The extended waiting period also makes it more difficult to determine the optimal scan timing, thereby increasing the risk of motion artifacts and variable image quality.

Dev: Current approaches still face two major limitations: limited flexibility in handling incomplete inputs and insufficient modeling of sequential enhancement dynamics across MRI phases.

Rosa: Network Architecture The proposed network consists of two core modules: the Enhanced Region-Guided Encoder (ERGE) and the Dynamic Feature Unification Module (DFUM).

Paper summary: Dev: So, to wrap up this discussion on the "Triple-Phase Sequential Fusion Network with Contrast Enhancement Kinetics for Hepatobiliary-Phase Liver MRI Synthesis," we're looking at a model that uses sequential information from pre-HBP sequences to synthesize HBP images <ref:2604.22904#pg1>.

Rosa: It claims this network generates images that are both visually realistic and physiologically plausible because it explicitly captures the temporal enhancement dynamics across different MRI phases <ref:2604.22904#pg3>.

Taro: The authors show that this model achieves contrast-to-noise ratios comparable to real HBP images, with a P value of zero point two six nine nine, and blinded evaluations confirmed they were visually comparable <ref:2604.22904#pg3>.

Dev: This suggests that TriPF-Net offers a way to generate HBP images that don't require the prolonged post-contrast delays we usually have to wait for <ref:2604.22904#pg1>.

Rosa: The paper shows that this approach is flexible enough to be robust against missing arterial or venous phase inputs, even though they still find that having all three pre-HBP phases helps them the most <ref:2604.22904#pg1>.

Taro: Ultimately, it’s a framework that tries to reduce reliance on delayed acquisition by using sequential information and patient clinical context to build more realistic HBP images <ref:2604.22904#pg3>.

Conclusion: Rosa: So, we've been talking about this paper called "Triple-Phase Sequential Fusion Network with Contrast Enhancement Kinetics for Hepatobiliary-Phase Liver MRI Synthesis," and now we need to talk about what that actually means for us.

Dev: Right, it’s trying to take the hassle out of getting those delayed hepatobiliary phase scans by using sequential information from earlier scans instead.

Rosa: Exactly, it's a way to build a full picture without making the patient wait around for a long time after they get their initial scan.

Taro: The authors are focusing on modeling how the enhancement changes over time across different MRI phases, which is actually pretty smart because those changes tell us something about how things are happening inside the liver tissue.

Dev: And they did this by using a network with two main parts, an encoder that looks at features from different points in the sequence and another part that fuses them together dynamically.

Rosa: They even made it more specific by adding patient stuff like age and bilirubin to guide how those images are generated, so it’s not just a generic reconstruction.

Taro: The results show they got lesion contrast that was comparable to the real hepatobiliary phase images, which is what doctors really want when they're looking for cancer.

Dev: But you have to remember the caveat: this whole system relies on having those three initial phases—arterial, portal venous, and then the HBP—to work well.

Rosa: And that’s where we get to think about what happens when the world doesn't cooperate and one of those inputs is missing or bad.

Taro: That’s our next big question for this paper: how does this network handle a situation where the input data isn't perfect?

Qiuli Wangb, Xinhuan Sunb, Fengxi Chena, Yongxu Liub, Jie Chenga, Lin Chena, Jiafei Chena, Yue Zhangd, Xiaoming Lia, Wei Chena

7T Magnetic Resonance Translational Medicine Research Center · Yu-Yue Pathology Research Center · Department of Clinical Laboratory Medicine · Laboratory of Intelligent Collaborative Computing University of Electronic Science and Technology of China

eess.IV, cs.CV

Submitted: 2026-04-24

Updated: 2026-11-09

Comments: 10 tables, 7 figures

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 88/100

The gist: Triple-Phase Sequential Fusion Network (TriPF-Net) is proposed to synthesize hepatobiliary phase (HBP) images by leveraging sequential information from pre-HBP sequences to reduce reliance on…

Key concepts

Enhanced Region-Guided Encoder (ERGE)
This module captures multiscale features from sequential MRI inputs using automatically generated masks for the liver and tumor regions. It extracts detailed information about focal lesions, liver tissue, and the surrounding abdominal area to guide the synthesis process.
Region-Guided Sequential Fusion Loss (LRGSF)
This loss function explicitly models how signal intensity changes over time across different MRI phases. It constrains the temporal evolution of background tissue, liver parenchyma, and tumor lesions separately based on their expected enhancement patterns in sequential scans.
Clinical Feature Encoding Module
This component takes patient clinical data like age, sex, bilirubin, and albumin. It maps this 22-dimensional clinical information into spatial feature maps within the network. This allows the synthesized image generation to be informed by the patient's specific physiological context.
Sequential Dynamics Modeling
The network explicitly learns and mimics the temporal signal-intensity evolution seen when contrast is administered over different MRI phases. By constraining these dynamics, TriPF-Net ensures that synthesized images reflect realistic enhancement kinetics across the sequence.

Terminology

Summary

Triple-Phase Sequential Fusion Network (TriPF-Net) is proposed to synthesize hepatobiliary phase (HBP) images by leveraging sequential information from pre-HBP sequences to reduce reliance on prolonged post-contrast delays

The Problem Addressed

Gadoxetate disodium-enhanced MRI is essential for the detection and characterization of hepatocellular carcinoma, but acquiring the hepatobiliary phase (HBP) requires a prolonged post-contrast delay which reduces workflow efficiency and increases the risk of motion artifacts. The extended waiting period also makes it more difficult to determine the optimal scan timing, thereby increasing the risk of motion artifacts and variable image quality. p. 0 Current approaches still face two major limitations: limited flexibility in handling incomplete inputs and insufficient modeling of sequential enhancement dynamics across MRI phases. p. 1

Network Architecture

The proposed network consists of two core modules: the Enhanced Region-Guided Encoder (ERGE) and the Dynamic Feature Unification Module (DFUM). The synthesis process maps the inputs to the synthesized HBP image, where wT1 is fixed to 1 while wAP and wVP are binary weights determined by the training stage. p. 3 The ERGE module captures multiscale sequential features across focal lesions, liver parenchyma, and global abdominal context using automatically generated masks for liver (ML) and tumor (MT) regions. p. 3

Sequential Dynamics Modeling

The network explicitly models the sequential enhancement dynamics across these three phases. This is achieved through the Region-Guided Sequential Fusion Loss (LRGSF), which mimics the temporal signal-intensity evolution observed across sequential MRI phases. p. 4 LRGSF introduces a clinically motivated constraint by separately constraining background (RB), liver parenchyma (RL), and tumor lesions (RT) according to their expected temporal behaviors. p. 4

Clinical Context Integration

TriPF-Net provides a multi-dimensional physiological context by integrating key clinical variables such as age, sex, total bilirubin, and albumin into a Clinical Feature Encoding Module. This module maps the patient’s 22-dimensional clinical data into spatial feature maps to effectively utilize the patient’s clinical prior information during image generation. p. 3

Performance and Validation

TriPF-Net achieved superior performance on datasets from two centers, reaching an MAE of 10.65 on the internal dataset and an MAE of 12. p. 041 on the external validation dataset. p. 0 Lesion-level evaluation showed that synthesized images achieved contrast comparable to real HBP images with no significant difference in CNR (P = 0. p. 02699). Blinded evaluations confirmed synthesized images were visually comparable to real HBP, suggesting potential for eliminating the need for delayed HBP acquisition. p. 0 Ablation studies showed that incorporating the Gabor filtering layer yielded the most significant performance gain, reducing MAE to 10. p. 76582.

Conclusion

In this study, TriPF-Net provides a flexible and physiologically informed framework for synthesizing hepatobiliary phase (HBP) MRI in patients with hepatocellular carcinoma. The proposed model generates HBP images that are both visually realistic and physiologically plausible by explicitly capturing sequential enhancement dynamics while maintaining robustness against missing arterial- or venousphase inputs. p. 9 Comprehensive validation demonstrated that TriPF-Net consistently outperforms state-of-the-art methods in both global reconstruction and lesion-level contrast, supporting its potential to streamline clinical workflows. p. 9 The findings suggest that TriPF-Net can generate HBP images with lesion contrast comparable to real HBP scans, supporting the potential to streamline clinical workflows. p. 8 The synthesized images achieve lesion-level contrast comparable to real HBP scans, supporting the potential to streamline clinical workflows. p. 8 p. 8 The synthesized images achieve lesion-level contrast comparable to real HBP scans, supporting the potential to streamline clinical workflows.

REFERENCES

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Enhanced flexibility in input handling allows TriPF-Net to synthesize HBP images even under the stochastic absence of one or both dynamic contrast-enhanced sequences. This means the system can generate high-fidelity HBP images using only T1-weighted imaging, AP and VP sequences, or even just T1, AP and HBP if VP is missing.

  2. Modeling sequential dynamics ensures physiological consistency by explicitly modeling the temporal signalintensity evolution observed across sequential MRI phases. This allows the AI to enforce biologically plausible transitions from pre-contrast T1 through vascular enhancement (AP/VP) to HBP, which improves the fidelity of synthesized images.

  3. Integration of clinical variables enhances plausibility by jointly integrating triple-phase imaging data with key clinical variables—including age, sex, and liver function biomarkers (serum bilirubin and albumin). This enables the model to generate HBP images that are not only visually realistic but also physiologically plausible, reflecting complex contrast agent excretion kinetics.

  4. The Region-Guided Sequential Fusion Loss explicitly models region-specific temporal behaviors by constraining the loss as LRGSF = lambdaBLback + lambdaLLliver + lambdaTLtumor. This guides the network to preserve the expected increase in hepatocellular uptake-related intensity for liver parenchyma and enforces lesion-to-liver contrast evolution for tumors, leading to improved lesion conspicuity.

  5. The inclusion of a Gabor filtering layer significantly boosts performance by yielding the most significant gain, reducing MAE to 10.6582 and increasing PSNR and SSIM substantially compared to the baseline model. This suggests that incorporating fine textural details captured by Gabor kernels is crucial for accurately modeling tissue-specific contrast uptake patterns.

Sources

Related papers