Proud to share that research co-authored by PurpleAI CTO Hoon Na has been published in Scientific Reports — and it tackles one of the hardest problems in stroke imaging: catching the strokes that are easiest to miss.
On non-contrast CT, small, non-LVO infarcts often look just like old white-matter change. Yet NCCT is still the first scan most stroke patients get — it's fast, low-cost, and available almost everywhere. That gap between what's scanned and what's seen is where patients fall through.
Newly published in Scientific Reports (Nature Portfolio), this pivotal, multi-reader, crossover, randomized study evaluated a deep learning algorithm for detecting acute ischemic stroke on NCCT — deliberately enriched for small infarcts. The work originated within SK Inc. C&C before PurpleAI's spin-off, and PurpleAI's Hoon Na is among its authors.
What the study found across 917 cases and 9 readers:
Standalone model performance: AUC 0.8144 (73.8% accuracy, 75.8% sensitivity, 72.6% specificity)
AI-assisted reading beat unassisted reading: 75.63% vs. 72.03% accuracy (+3.60%, p < 0.001)
Biggest gain for non-radiologist physicians: +5.38% (75.35% vs. 69.97%, p < 0.001)
The signal that matters most: the readers who gained the most were the least specialized. That's not just a benchmark win — it points toward more timely and more equitable stroke diagnosis in exactly the settings where neuroradiology expertise isn't always at the bedside.
Congratulations to the full author team: Tae Jin Yun, Jinwook Choi, Hoon Na — and to our collaborators at Seoul National University Hospital and Ajou University Medical Center.
👇 Full open-access paper linked in the comments.
#StrokeCare #NCCT #RadiologyAI #ClinicalAI #PurpleAI