Transformer-Enhanced Generative Adversarial Networks for Improving MR Image Quality in Prostate Imaging

Jie Bao, Litao Zhao, Ying Hou, Yueting Su, Libiao Ji, Junkang Shen, Ximing Wang, Hailin Shen, Pheng‐Ann Heng, Chunhong Hu, Yudong ZhangResearch Square 2025 · DOI

Abstract

Consistently high-quality and standardized prostate MRI is crucial for reliable PI-RADS interpretation and accurate prostate cancer detection, particularly given the heterogeneity in acquisition protocols across institutions. In this study, we develop and validate an innovative anatomy-aware pyramid-spatially adaptive normalization and transformer-enhanced CycleGAN (PST-CyGAN) framework for standardizing and enhancing prostate MRI quality in large-scale multicenter datasets. This study included 2,207 patients for model training and 497 for validation from 8 centers. PST-CyGAN was built upon CycleGAN with integrated anatomical attention, pyramid-spatially adaptive normalization, and transformer enhancement. High-resolution T2-weighted imaging (HR-T2WI) with quality control was used to ensure the model’s ability to generate high-quality images. PST-CyGAN consistently outperformed state-of-the-art (SOTA) models across all metrics, significantly upgrading PI-QUAL scores, especially in initially low-quality scans.

Introduction

Prostate MRI serves as the cornerstone for non-invasive prostate cancer detection and management. However, the clinical utility of these scans is often compromised by significant inter-institutional variability. Differences in hardware, software versions, and acquisition protocols lead to inconsistent image quality, manifesting as varying levels of noise, resolution, and artifacts. Such inconsistencies make the standardized application of the Prostate Imaging Reporting and Data System (PI-RADS) challenging for radiologists. While deep learning, particularly Generative Adversarial Networks (GANs), has shown promise in image enhancement, traditional architectures often suffer from "hallucinations" or fail to maintain global anatomical consistency. To overcome these limitations, we present PST-CyGAN, a framework designed to provide standardized, high-quality synthetic T2-weighted images (T2WI) across diverse clinical environments.

Methods

The proposed PST-CyGAN (Anatomy-aware Pyramid-Spatially adaptive normalization and Transformer-enhanced CycleGAN) architecture introduces three primary innovations to the standard CycleGAN framework:

  1. Spatially-Adaptive Normalization (SPADE): By incorporating anatomical masks during the generation process, the model is forced to respect the underlying structure of the prostate gland. This "anatomical awareness" acts as a structural anchor, preventing the geometric distortions often seen in unsupervised image translation.
  2. Pyramid Module: This component allows the network to process information at multiple spatial scales simultaneously. This multi-scale approach is essential for accurately capturing both the broad location of the prostate and the fine-grained details of its internal architecture.
  3. Transformer Enhancement: Unlike traditional convolutional layers that focus on local pixel neighborhoods, the integrated Transformer module utilizes self-attention mechanisms to capture long-range spatial dependencies. This significantly improves the perceptual quality and global coherence of the enhanced images.

The study utilized a large-scale multicenter dataset comprising 2,704 patients. The training phase involved 2,207 patients, including 941 who received high-resolution T2WI (HR-T2WI) to serve as the target domain. A critical aspect of our methodology was the implementation of a "quality-controlled" training strategy, where the model was trained exclusively on the highest-quality, radiologist-curated scans. This approach ensures the model learns a mapping toward a "gold-standard" visual representation rather than merely averaging across varying input qualities.

Results

PST-CyGAN demonstrated superior performance in both quantitative and qualitative evaluations. In ablation studies, the model consistently outperformed state-of-the-art (SOTA) benchmarks, including standard CycleGAN, S-CyGAN, and PS-CyGAN. Quantitative metrics showed significant improvements:

  • Peak Signal-to-Noise Ratio (PSNR): 29.3 ± 0.5
  • Structural Similarity Index (SSIM): 0.558 ± 0.116
  • Feature Correlation Score (FCS): 0.915 ± 0.028

Clinical validation involved five blinded radiologists who assessed the generated images using the 10-point PI-QUAL scoring system. The quality-controlled PST-CyGAN model upgraded the PI-QUAL scores in 38% of cases, with the most pronounced benefits observed in scans that were initially of poor quality (PI-QUAL < 6). Conversely, scans that were already of high quality saw minimal changes, demonstrating an "inverse benefit" where the model provides the most assistance where it is needed most. Prospective testing on external datasets with varying voxel sizes further confirmed the model's robust generalizability, with quality improvements ranging from 20% to 29%.

Discussion and Conclusion

The PST-CyGAN framework represents a significant advancement in the standardization of prostate MRI. By combining anatomical anchors with transformer-based global context, the model effectively bridges the gap between heterogeneous acquisition protocols. Our findings underscore the importance of training data quality; the model's ability to generate "gold-standard" images was directly tied to the use of curated, high-resolution training targets.

Despite these successes, certain limitations remain. The current model is optimized for T2-weighted imaging; extending this framework to functional sequences like Diffusion-Weighted Imaging (DWI) and Apparent Diffusion Coefficient (ADC) maps is a necessary next step for comprehensive multiparametric MRI enhancement. Additionally, future iterations could benefit from adaptive processing to avoid unnecessary alterations to already high-quality scans. In conclusion, PST-CyGAN offers a robust solution for enhancing and standardizing prostate MRI, potentially reducing diagnostic variability and improving the accuracy of prostate cancer detection in clinical practice.

Summary updated

Abstract

<title>Abstract</title> Consistently high-quality and standardized prostate MRI is crucial for reliable PI-RADS interpretation and accurate prostate cancer detection, particularly given the heterogeneity in acquisition protocols across institutions. In this study, we develop and validate an innovative anatomy-aware pyramid-spatially adaptive normalization and transformer-enhanced CycleGAN (PST-CyGAN) framework for standardizing and enhancing prostate MRI quality in large-scale multicenter datasets. This study included 2,207 patients for model training and 497 for validation from 8 centers, all of whom underwent prostate MRI between January 2017 and May 2025. PST-CyGAN was built upon CycleGAN with integrated anatomical attention, pyramid-spatially adaptive normalization, and transformer enhancement. High-resolution T2-weighted imaging (HR-T2WI; resolution: 0.35 × 0.35 × 2.0 mm) with quality control was acquired from 941 patients to ensure the model’s ability to generate high-quality images. Ablation tests were conducted comparing PST-CyGAN against three state-of-the-art (SOTA) models: CycleGAN, S-CyGAN, and PS-CyGAN. For clinical validation, five blinded radiologists evaluated the image quality of PST-CyGAN-generated HR-T2WI using a 10-point PI-QUAL scoring system. Changes in PI-QUAL scores compared to standard-resolution (SR) T2WI were analyzed using Bland–Altman plots and paired Wilcoxon tests. In ablation testing, PST-CyGAN consistently outperformed all three SOTA models across all metrics: PSNR (29.3 ± 0.5), SSIM (0.558 ± 0.116), FCS (0.915 ± 0.028), and LPIPS (0.169 ± 0.056) (all <italic>P</italic> &lt; 0.01). Improvements were most notable in initially low-quality scans (PI-QUAL &lt; 6). In multi-reader validation, PST-CyGAN trained with quality-controlled data achieved the highest clinical benefit, upgrading PI-QUAL scores in 38% of cases and downgrading in 20%, compared to 11% upgrades and 69% downgrades without quality control ( <italic>P</italic> &lt; 0.01). Prospective testing demonstrated PI-QUAL score improvements of 25.5%, 20.0%, and 29.1% for voxel sizes of 1.5×1.5×3.3 mm, 0.7×0.7×3.3 mm, and 0.4×0.4×2.0 mm, respectively, confirming robust generalizability. PST-CyGAN significantly and reproducibly enhances prostate T <sub>2</sub> WI quality across diverse centers and acquisition protocols, exceeding all SOTA models on quantitative metrics. By generating standardized, high-resolution synthetic T <sub>2</sub> WI—especially from initially low-quality scans—it reduces inter-site variability, supports consistent PI-RADS application, and may improve prostate MRI interpretation in clinical workflows.
BibTeX
@inproceedings{bao2025transformer,
  author = {Jie Bao and Litao Zhao and Ying Hou and Yueting Su and Libiao Ji and Junkang Shen and Ximing Wang and Hailin Shen and Pheng‐Ann Heng and Chunhong Hu and Yudong Zhang},
  title = {{Transformer-Enhanced Generative Adversarial Networks for Improving MR Image Quality in Prostate Imaging}},
  booktitle = {Research Square},
  year = {2025},
  doi = {10.21203/rs.3.rs-7833508/v1},
}
Figure