Answers are grounded in this paper and cite their sources.
Ischemic stroke is a leading cause of disability and mortality worldwide. Effective treatment is associated with fast and effective lesion quantification from image observations. NCCT is the common first imaging to discard other neurological conditions, but it has limited sensitivity for detecting early ischemic changes. In addition, DW-MRI studies are considered the gold standard for early lesion characterization through DWI sequences and ADC maps. Nonetheless, the analysis from MRI images is delayed, restricted for some patient conditions, and typically limited because of accessibility constraints. Although recent approaches have addressed these limitations by synthesizing DW-MRI from NCCT, they often fail to reproduce ischemic lesions. Moreover, their performance is usually reported with whole-image metrics, which can overestimate image quality. This work explores the capability of conditional diffusion models to synthesize DWI sequences and ADC maps from NCCT to support care protocols in early stroke scenarios. In particular, we explore both a direct synthesis from NCCT to ADC and a conditional cascade configuration. In the cascade, the NCCT is used as a conditioning input to generate synthetic DWI findings. Then, the corresponding ADC maps are generated from the synthetic DWI. The conditional diffusion models outperformed the U-Net, Pix2Pix and CycleGAN baselines in the brain region, reaching an SSIM of 0.53 for DWI and up to 0.36 for ADC. Direct diffusion achieved a higher global ADC fidelity than cascade diffusion (SSIM 0.87 vs. 0.83), whereas cascade diffusion obtained a lower ADC error within the lesion (MAE 0.19 vs. 0.22).
No takes yet. Share an insight, caveat, or question.
Ramírez et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: