RESEARCH ON ISCHEMIC STROKE MRI IMAGE CLASSIFICATION BASED ON CONVOLUTIONAL NEURAL NETWORK

Abstract

Stroke is a group of diseases that damage brain tissue caused by sudden local cerebral blood circulation disorder. If not diagnosed and treated in time, the patient will lose a large number of neurons and even die. Among the examination methods, using magnetic resonance scans is more accurate than CT images, but it also has other risk factors. In response to these existing problems, this paper attempts to use a stroke classification model based on pre-trained CNN networks, which can help doctors detect MRI images of ischemic stroke and reduce time delays. This paper uses two real datasets. Dataset 1 includes 1112 cases of ischemic stroke and 1202 normal MRI images, while dataset 2 includes 1008 cases of ischemic stroke and 1002 normal MRI images. After data processing and noise reduction, two processed datasets, which contain 4448 cases, 4808 cases and 4032 cases, 4008 cases of ischemic stroke and normal MRI images, respectively, were obtained. Features are obtained by extracting the last pooling layer and the last fully connected layer of the pre-trained CNN network ResNet50 and DenseNet201. After combining the features, the iterative mRMR feature selection method is used to select image features. A support vector machine classifier with 10-fold cross-validation is used to classify the data. Finally, IMV is applied to the prediction vectors to get the voting results, and the best results are selected from the results. The evaluation indicators of the final experimental results in dataset 1 yielded 94.74%, 93.56%, 96.22%, and 94.87% results for accuracy, recall, precision, and F1 score, respectively. The evaluation indicators in dataset 2 yielded 92.24%, 92.21%, 92.26%, and 92.23% results for accuracy, recall, precision, and F1 score, respectively. These results clearly demonstrate that ischemic stroke can be successfully classified based on MRI and confirm the success of the hybrid feature engineering approach.