An Ensemble Deep Learning Network in Classifying the Early CT Slices of Ischemic Stroke Patients

Abstract

The human brain is the body's most complicated organ. Constant blood flow is essential for the sustained functioning of the brain. A blocked blood vessel's interruption of blood supply prevents oxygen and nutrients to the brain tissues. This results in a life-threatening brain disease called Ischemic Stroke. Computed Tomography (CT) images are widely used in the diagnosis of Ischemic Stroke because of their faster acquisition and compatibility with most life support devices. CT acquired from the patients who arrived with stroke symptoms is Primary CT (PCT). After some hours CT taken for the same patient is Secondary CT (SCT). Stroke lesions may not be visible in PCT, whereas visible in SCT. Learning the features automatically using a Convolutional Neural Network (CNN) is essential to classify normal and abnormal CT slices. These networks are capable of learning the global features effectively for image classification. Though this CNN approach works, achieving desired accuracy was challenging. Different architectures considered for this CNN experimentation are VGG1, VGG2, VGG3, VGG16, InceptionV3, and ResNet50. This novel work provides a detailed explanation of the three experiments conducted using PCT and SCT slices. Three experiments are conducted using SCT and PCT slices. The pretrained VGG16, ResNet50, and InceptionV3 networks with the ImageNet database are applied as a first approach. Both SCT and PCT slices are used for testing alone. It resulted in 49.22%, 47.076% and 49.36% classification accuracy. In the second approach, different models were trained for classification from PCT and SCT slices. This includes the networks like VGG1, VGG2, VGG3, VGG16, VGG16 with dropout, ResNet, ResNet50 with lambda regularization, InceptionV3, and InceptionV3 with lambda regularization. The accuracies achieved are 68%, 69.4%, 72%, 78.2%, 79.1%, 77%, 77.8%, 79.6% and 80.1%. The accuracy was improved with dropout and lambda regularization. The networks with high accuracy are selected and an ensemble model is developed as a third approach. ResNet50, VGG16, and InceptionV3 are combined to form an ensemble network. This ensemble network yielded an accuracy of 81.98% when SCT and PCT slices are used for both training and testing. And produced 74% accuracy when PCT slices alone were used. Also produced 93.76% accuracy when SCT slices alone were used.