Dual Low-b-value-driven U-shaped fusion GAN for synthesizing high-b-value prostate DWI.
Authors
Affiliations (9)
Affiliations (9)
- Jiangsu Province Engineering Research Center of Smart Wearable and Rehabilitation Devices, School of Biomedical Engineering and Informatics, Nanjing Medical University, Nanjing, 211166, China.
- Department of Radiology, Shanghai Fifth People's Hospital, Fudan University, Shanghai, 200240, China.
- The Department of Radiology, The First Affiliated Hospital of Nanjing Medical University, Nanjing, 210029, China.
- Department of Radiology, Nanjing First Hospital, Nanjing Medical University, No.68, Changle Road, Nanjing, 210006, China.
- Department of Radiology, Zhongda Hospital Southeast University, Nanjing, 210009, China.
- Department of Medical Information Engineering, Kangda College of Nanjing Medical University, Lianyungang, 222000, China.
- The Department of Radiology, The First Affiliated Hospital of Nanjing Medical University, Nanjing, 210029, China. [email protected].
- Jiangsu Province Engineering Research Center of Smart Wearable and Rehabilitation Devices, School of Biomedical Engineering and Informatics, Nanjing Medical University, Nanjing, 211166, China. [email protected].
- Jiangsu Province Engineering Research Center of Smart Wearable and Rehabilitation Devices, School of Biomedical Engineering and Informatics, Nanjing Medical University, Nanjing, 211166, China. [email protected].
Abstract
High-b-value (b = 2000 s/mm²) diffusion-weighted imaging (DWI) is vital for prostate disease detection and characterization due to superior tumor-to-background contrast, but its direct acquisition is time-consuming, technically demanding, and prone to noise and artifacts, limiting routine clinical use. This study aims to synthesize high-b-value prostate DWI from low-b-value data via a novel deep learning method. A dual low-b-value-driven U-shaped Fusion Generative Adversarial Network (UsFGAN) was proposed, integrating three core components: (1) U-Net-based dedicated subnetworks (with skip connections) for feature extraction from two low-b-values (b = 50/1000 sec/mm²); (2) Swin-Transformer with residual blocks (STRB) to capture local/long-range pixel dependencies; (3) hierarchical fusion network with multiple feature fusion blocks (MFFB) for adaptive multi-scale feature combination. Validation was done on a multi-center dataset of 280 subjects (6440 DWI slices). The proposed method outperformed state-of-the-art models (CycleGAN, Pix2Pix, DiscoGAN): peak signal-to-noise ratio = 36.14 dB, structural similarity index = 0.91, LPIPS = 0.09, FID = 8.87. Synthesized high-b-value DWI achieved 86.3% accuracy in prostate lesion detection. Radiologist qualitative evaluation confirmed synthesized images were comparable to real high-b-value scans in noise suppression, artifact reduction, and diagnostic acceptability. UsFGAN effectively leverages dual low-b-value complementary information to synthesize high-quality high-b-value prostate DWI. It exhibits superior performance and clinical diagnostic value, promising to reduce scan time and improve prostate disease assessment.