CT Perfusion Map Generation from Multiphase CTA Using a Generative Adversarial Model for Acute Ischemic Stroke

Multiphase CT Angiography (mCTA) has shown potential as a diagnostic tool for acute ischemic stroke because it captures dynamic changes in the cerebral vasculature. However, mCTA has limitations in assessing brain tissue perfusion, which reduces its clinical interpretability. To address this limitation, we aimed to develop a generative adversarial network (GAN) that generates CTP-like maps from mCTA. This approach aims to improve the interpretability of mCTA.

A total of 714 cases with NCCT, CTP, mCTA, and follow-up NCCT/MRI were analyzed across internal and external data sets. A GAN was trained to generate multiparametric CTP maps (Tmax, CBF, CBV). The performance of the model was evaluated using the Structural Similarity Index (SSIM), peak signal-to-noise ratio (PSNR), and Fréchet Inception Distance (FID) compared with actual CTP maps. Clinical utility was assessed by predicting infarct core and penumbra using threshold-based segmentation and evaluating metrics such as the Dice coefficient, area under the receiver operating characteristic curve (AUC) of dichotomized infarct volumes of < 70 mL, and mismatch ratio following DEFUSE 3 criteria, compared with the ground truth of actual CTP prediction.

The GAN achieved SSIM, 0.65-0.66; PSNR, 20.4-20.8; and FID, 15.8-17.0 on internal data, surpassing both CycleGAN (SSIM: 0.608-0.642, PSNR: 18.2-19.7, FID: 27.6-32.5) and Pix2Pix (SSIM: 0.630-0.645, PSNR: 19.5-19.7, FID: 19.4-20.8) across all metrics. Predicted penumbra and infarct core showed Dice coefficients of 0.672 and 0.468, with strong correlations (penumbra: 0.921, core: 0.902) and AUCs of 0.854 (95% CI, 0.819-0.888) (mismatch ratio) and 0.850 (95% CI, 0.817-0.884) (dichotomized infarct core). External data validation yielded Dice coefficients of 0.481 (penumbra) and 0.301 (core) with AUCs of 0.720 (95% CI, 0.589-0.808) (mismatch ratio) and 0.703 (95% CI, 0.528-0.794) (dichotomized infarct core).

The GAN effectively generated CTP-like maps from mCTA, improving interpretability and demonstrating promising diagnostic performance, particularly for resource-limited settings.