ORIGINAL RESEARCH article
Front. Phys.
Sec. Medical Physics and Imaging
Volume 13 - 2025 | doi: 10.3389/fphy.2025.1661146
This article is part of the Research TopicMethods and Strategies for Integrating Medical Images Acquired from Distinct ModalitiesView all 7 articles
High-Fidelity Medical Image Generation: Controllable Synthesis of High-Resolution Medical Images via Hierarchical Fusion in Vector-Quantized Generative Networks
Provisionally accepted
Guangfa Tang1,2,3
Shanshan Cai4
Xiangjun Meng5Siyan Huo6Mengbo Wang7Zichen Lu3Zhuokang Chen5Xiaoling Luo2*- 1Xinhua College of Sun Yat-sen University, Guangzhou, China
- 2Shenzhen University, Shenzhen, China
- 3Dongguan Zhongke Institute of Cloud Computing, Dongguan, China
- 4Lancaster University, Lancaster, United Kingdom
- 5Dongguan People's Hospital, Dongguan, China
- 6Medical College of Wisconsin, Milwaukee, United States
- 7Guangzhou Xinhua University, Guangzhou, China
Select one of your emails
You have multiple emails registered with Frontiers:
Notify me on publication
Please enter your email address:
If you already have an account, please login
You don't have a Frontiers account ? You can register here
Objective: High-resolution medical images are scarce. Existing methods perform poorly at high resolutions, struggling with small lesions, detail loss, anatomical distortion, computational cost, and mode collapse. We develop a novel framework addressing these challenges.Methods: Using 255 X-rays and 1,657 public lung CTs, we propose a two-stage synthesis: 1) SinGAN generates lesions preserving texture, 2) HiResMed-VQGAN (with HDFB module) synthesizes backgrounds. Lesions are fused at anatomically plausible locations. Compared to DDM/StyleSwin/VQGAN/SinGAN using FID/LPIPS/PSNR/SSIM. Clinicians performed Visual Turing Tests.Results: Our method reduced FID by 43.3%, LPIPS by 5%, increased PSNR by 4% and SSIM by 6% versus VQGAN, with 83% faster convergence. Clinicians misclassified 55% of synthetic images as real.Conclusion: The framework generates high-resolution medical images with controlled small lesions while preserving anatomical fidelity and computational efficiency.
Keywords: Controllable synthesize, two routes synthesis strategy, high-resolution medicalimage generation, hierarchical dual-path learning, Detail preservation, High fidelity
Received: 09 Jul 2025; Accepted: 29 Aug 2025.
Copyright: © 2025 Tang, Cai, Meng, Huo, Wang, Lu, Chen and Luo. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
* Correspondence: Xiaoling Luo, Shenzhen University, Shenzhen, China
Disclaimer: All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.