Comparison support vector machines and K-nearest neighbors in classifying Ischemic stroke by using convolutional neural networks as a feature extraction

The paper introduces the hybrid method of Convolutional Neural Network (CNN) and machine learning methods as a classifier, that is Support Vector Machines and K-Nearest Neighbors for classifying the ischemic stroke based on CT scan images. CNN is used as a feature extraction and the machine learning methods used to replace the fully connected layers in CNN. The proposed method is used to reduce computation time and improve accuracy in classifying image data, because we know that deep learning is not efficient for small amounts of data, where the data we use is only 93 CT scan images obtained from Cipto Mangunkusumo General Hospital (RSCM), Indonesia. The architecture of CNN used in this research consists of 5 layers convolutional layers, ReLU, MaxPooling, batch normalization and dropout. The elapsed time required for CNN is 7.631490 seconds. The output of feature extraction is used as an input for SVM and KNN. SVM with linear kernel can correctly classify ischemic stroke, with 100% accuracy in the training model and 96% accuracy in testing model with a test size of 60%. KNN classify ischemic stroke, with 97.3% (#neighbors = 5) accuracy in training model with a test size of 60% and 90% (#neighbors = 10, 15, 25) accuracy in the testing model with a test size of 10%. Based on these results, the SVM produces the higher accuracy compared to KNN in classifying ischemic stroke using CNN as feature extraction based on CT scan images with a computation time of only 8.0973 seconds.