Deep Learning-Based Gadolinium-Free Virtual Contrast-Enhanced MRI for Predicting Immediate HIFU Ablation Outcomes in Uterine Fibroids.
Authors
Affiliations (3)
Affiliations (3)
- Department of Radiology, The Third People's Hospital of Chengdu, Chengdu, China (C.L., H.L., J.L., K.L., X.P., J.G.).
- Department of Radiology, The First Affiliated Hospital of Chongqing Medical University, Chongqing, China (Z.X.).
- Department of Radiology, The Third People's Hospital of Chengdu, Chengdu, China (C.L., H.L., J.L., K.L., X.P., J.G.). Electronic address: [email protected].
Abstract
Accurate assessment of microvascular perfusion is essential for predicting high-intensity focused ultrasound (HIFU) ablation outcomes in uterine fibroids. Contrast-enhanced T1-weighted imaging (CE-T1WI) is widely used for vascular evaluation; however, gadolinium-based agents pose safety concerns and may interfere with temperature monitoring. This study aimed to generate synthetic CE-T1WI (Syn-CE-T1WI) from T2-weighted imaging (T2WI) using deep learning for predicting immediate HIFU ablation outcomes. A mask-guided 3D conditional generative adversarial network (3D-cGAN) was developed using magnetic resonance imaging data from 300 patients (240 training, 60 testing). Syn-CE-T1WI was generated from T2WI. Image quality was assessed using peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and mean absolute error (MAE). Three radiologists performed blinded evaluations of image authenticity and quality. A binary classification task was used to compare Syn-CE-T1WI and Real-CE-T1WI for predicting HIFU outcomes. In the test set, PSNR was 27.05 ± 2.21 dB, SSIM was 0.825 ± 0.12, and MAE was 0.023. Radiologists achieved a mean accuracy of 52.8% in distinguishing real from synthetic images (all P > 0.05). Overall, 88.9% of Syn-CE-T1WI were rated clinically acceptable. For outcome prediction, Syn-CE-T1WI achieved areas under the receiver operating characteristic curve (AUCs) of 0.713-0.763, comparable to Real-CE-T1WI (0.725-0.775; all P > 0.05). Combining Syn-CE-T1WI with T2WI yielded AUCs of 0.775-0.800, numerically higher than T2WI alone (0.713-0.763), and similar to T2WI + Real-CE-T1WI (0.800-0.813; all P > 0.05). The mask-guided 3D-cGAN enabled the generation of high-fidelity, gadolinium-free Syn-CE-T1WI that may support noninvasive prediction of immediate HIFU ablation outcomes in uterine fibroids.