Accurate multi-b-value DWI generation using two-stage deep learning: multicenter study.
Authors
Affiliations (5)
Affiliations (5)
- Department of Radiology, Sir Run Run Hospital, Nanjing Medical University, 109 Longmian Road, Nanjing, Jiangsu 211002, People's Republic of China.
- Department of Radiology, Ma'anshan People's Hospital, 45 Hubei Road, Maanshan, Anhui, People's Republic of China.
- School of Literature, Nanjing Normal University, 122 Ninghai Road, Nanjing, Jiangsu 210097, People's Republic of China.
- Department of Radiology, Air Force Medical Center, Air Force Medical University, 30 Fucheng Road, Haidian District, Beijing 100142, People's Republic of China.
- Department of Radiology, Sir Run Run Hospital, Nanjing Medical University, 109 Longmian Road, Nanjing, Jiangsu 211002, People's Republic of China. Electronic address: [email protected].
Abstract
This study aimed to develop and validate a two-stage deep learning framework (DC<sup>2</sup>Anet-MineGAN) for high-fidelity multi-organ, multi-b-value DWI synthesis and accurate ADC restoration, addressing real-world limitations of DWI acquisition. This retrospective study included DWI images collected from three hospitals (from March 2020 to March 2025) and the TCIA database, with a total of 50,000 images across five anatomical regions (brain, breast, abdomen, neck, pelvis) and b-values (0-1000 s/mm<sup>2</sup>). Images were split 8:2 into training and test sets. The two-stage model used DC<sup>2</sup>Anet for low-to-high b-value synthesis per organ and MineGAN for interpolation to arbitrary b-values. Performance was evaluated using MSE, MAE, PSNR, and SSIM, and radiologist Likert ratings with ICC for synthetic images at non-trained b-values. Group differences were assessed statistically. Of 50,000 images, synthetic ADC values closely matched ground truth across all regions (mean difference < 0.02; all p > 0.05), with SSIM > 0.81 and PSNR > 74 for all b-values. For unseen b-values (b = 75, 850), 90/120 (75 %) and 60/120 (50 %) images were scored as excellent by radiologists; ICC exceeded 0.92 for both b-values. No significant difference was observed between synthetic and real ADC, and no regions showed inferior image quality. Recognizing potential hallucination or distortion, further multi-center clinical validation is required. The proposed two-stage DC<sup>2</sup>Anet-MineGAN framework achieves accurate, high-quality DWI synthesis and ADC reproduction across multiple b-values and anatomical regions, overcoming clinical DWI limitations and supporting reliable quantitative imaging.