Comparative Analysis of Deep Neural Network and Texture-Based Classifiers for Recognition of Acute Stroke using Non-Contrast CT Images

This work presents a computer technology for automatic recognition of acute stroke using non-contrast computed tomography brain images. The early diagnosis of acute stroke is of primary importance for deciding on a method for further treatment, and the developed system aims at assisting a radiology specialist in the decision making process. We consider deep neural network and texture-based classifiers in order to compare their efficiency on a moderate-sized sample of patients with acute stroke. We use U-net as a basic architecture of the neural network, and Haralick textural features, extracted from images, for kNN, SVM, Random Forest and Adaboost classifiers. Experiments with real CT images using cross-validation technique show that deep neural network outperforms the considered texture-based classifiers; however, the latter are faster in training. We demonstrate that texture-based approach is able to give potentially useful additional information for stroke recognition, such as estimates of textural features importance; visualization of differences in positive and negative class distributions.