A Multi-modal Deep Learning Framework for Final Infarct Prediction in Acute Ischemic Stroke: Combining CTA, NCCT, and Clinical Data
Abstract: Timely treatment decisions facilitate optimal outcomes in acute ischemic stroke patients, wherein computed tomography (CT) imaging data represents the primary imaging modality used for initial clinical assessment. This study presents and evaluates a novel deep learning architecture that integrates non-contrast computed tomography (NCCT) images, CT angiography (CTA) images, and mean and maximum projections of the CTA images, along with clinical data for tissue outcome predictions in patients with acute ischemic stroke. The proposed model is based on a convolutional neural network (CNN), including an encoder module, multiple fully connected layers, and a decoder. We trained and evaluated the proposed deep learning model using data from the ISLES 2024 challenge, which included 150 patients, of which we allocated 80% for training and 20% for evaluation. The evaluation showed that our model achieved a mean Dice score of 0.04 (\(\pm 0.07\)), a mean absolute volume difference of 35 ml (\(\pm 24.7\) ml), a mean lesion count difference of 207 (\(\pm 28.65\)), and a mean lesion-wise F1 score of 0.13 (\(\pm 0.15\)). While our model overestimated the size of small infarct cores, its satisfactory performance detecting larger lesions highlights the potential of simple and widely-available CTA images and its mean and maximum projections, along with the NCCT data and relevant clinical information for predicting final infarct brain tissue in patients with acute ischemic stroke.
Loading