Neurologists-level interpretable CT-based deep neural network for prediction of hemorrhagic transformation after ischemic stroke

Abstract

Background Hemorrhagic transformation (HT) is a severe complication following acute ischemic stroke, associated with neurological deterioration and poor clinical outcomes. Deep learning represents a promising tool for HT prediction. Methods We conducted a retrospective analysis of 474 acute ischemic stroke cases (231 HT and 243 non-HT) admitted to Beijing Tiantan Hospital from April 2014 to November 2022. We constructed a dataset from this cohort and randomly partitioned it into training and validation sets. Subsequently, we developed a model utilizing convolutional neural networks (CNNs) and residual networks based on computed tomography (CT) scans to predict HT after ischemic stroke. Results The final dataset consisted of 613 CT scans. The model achieved an F1 score of 78.94% (95% CI, 67.7–86.4). The Area Under the Curve (AUC) was 0.842 (95% CI, 75.8–92.1), sensitivity was 71.55% (95% CI, 60.6%−85.0%), and accuracy was 74.52% (95% CI, 63.9%−83.2%). Conclusion By combining plain CT scans with deep learning methodologies, we developed a clinically applicable model with demonstrable interpretability. Primarily designed to predict HT after acute ischemic stroke, this model demonstrated significant performance advantages in testing compared to both clinical physicians and similar existing models.

Ähnliche Arbeiten