Rapid Assessment of Acute Ischemic Stroke by Computed Tomography Using Deep Convolutional Neural Networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Acute stroke is a leading cause of death and disability globally.
• Rapid and accurate diagnosis is critical for effective acute stroke treatment.
• Current diagnostic methods require timely interpretation by specialists.

Purpose of the Study:

• To propose an automatic identification scheme for acute ischemic stroke using deep convolutional neural networks (DCNNs).
• To evaluate the performance of different DCNN architectures (AlexNet, Inception-v3, ResNet-101) for stroke detection.
• To assess the utility of transfer learning in improving DCNN performance with limited data.

Main Methods:

• A dataset of 1254 non-contrast computed tomographic (NCCT) images from 96 patients with acute ischemic stroke and 681 normal controls was used.
• Deep convolutional neural networks (DCNNs), including AlexNet, Inception-v3, and ResNet-101, were trained and evaluated.
• Transfer learning with ImageNet parameters was employed to augment training on limited data, with models validated using tenfold cross-validation and an independent dataset.

Main Results:

• AlexNet trained from scratch achieved 97.12% accuracy, 98.11% sensitivity, and 96.08% specificity on the primary dataset.
• Transfer learning models (AlexNet, Inception-v3, ResNet-101) achieved accuracies between 90.49% and 95.49% on the primary dataset.
• On an independent dataset, transferred AlexNet, Inception-v3, and ResNet-101 achieved accuracies of 81.77%, 85.78%, and 80.89%, respectively, outperforming AlexNet trained from scratch (60.89%).

Conclusions:

• The proposed DCNN architecture functions as a computer-aided diagnosis system for acute ischemic stroke.
• Training DCNNs from scratch can yield customized models for specific scanners.
• Transfer learning generates more generalized models, offering valuable diagnostic suggestions to radiologists for acute ischemic stroke detection.