Quantitative generation of diffusion-weighted imaging from non-contrast CT in acute ischemic stroke: a multi-scanner deep-learning model with external validation.

Background: Acute ischemic stroke (AIS) is a leading cause of death and disability worldwide, where early diagnosis and treatment are crucial for effective management. Computed tomography (CT), particularly non-contrast CT, is widely used due to its speed and low cost; however, it has limited sensitivity and specificity. Although diffusion-weighted imaging (DWI) provides higher diagnostic accuracy, it is time-consuming and may delay treatment. This study aims to investigate the feasibility of using deep learning techniques to convert non-contrast CT images of patients with cerebral ischemic stroke into DWI images. Methods: The dataset used in this study consisted of non-contrast CT images and DWI from 224 patients between May 2013 and April 2022, and it was divided into training (n=180), validation (n=20), and testing (n=24) cohorts. The proposed methodology used a modified ControlNet, a deep learning architecture known for its ability to learn complex mappings, to learn the intricate relationship between the two modalities and generate synthetic DWI based on non-contrast CT scans of patients with acute ischemic stroke. Model performance was evaluated using peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and normalized mean squared error (NMSE), together with Bland-Altman analysis, voxel-wise area under the receiver operating characteristic curve (AUC) for stroke core discrimination, and radiologist-based qualitative scoring. Results: On the independent testing cohort, the proposed model achieved a PSNR of 29.32±1.84 dB, an SSIM of 0.867±0.032, and an NMSE of 0.293±0.041, outperforming U-Net-, Pix2Pix-, CycleGAN-, and DualGAN-based comparison methods in quantitative and visual evaluations. For stroke core discrimination based on synthesized DWI, the proposed method achieved the highest voxel-wise AUC (0.764±0.026) among all compared methods. In case-level error analysis, the model showed 1 false-positive (FP) case (4.2%) and 4 false-negative (FN) cases (16.7%) in the 24 test cases. Qualitative evaluation by four radiologists showed high image quality, with a mean overall score of 4.78/5. The average inference time was 2.46 s per slice, corresponding to approximately 5-7 min for a full brain volume. Conclusions: The conversion of stroke CT scans into DWI using our methodology is not just feasible, but it has the potential to revolutionize bedside decision-making. This suggests a broader potential application to transform diagnostic imaging, particularly in the context of stroke diagnosis. Our method provides clinicians with valuable information, enabling the production of advanced DWI-like images on demand in emergency settings.