Local Mean Gradient Pattern (LMGP): A novel approach made for the classification of brain CT scan images

Abstract

ObjectiveThe visual descriptor methods like Local Binary Pattern (LBP) capture anatomical structures in captured images along with their disparities, which can be exploited by suitable methods for the diagnosis of medical anomalies. We developed a Local Mean Gradient Pattern (LMGP), based partly on LBP, a feature extraction algorithm for the classification of Computed Tomography (CT) images of the brain into normal, ischemic, or hemorrhage categories. MethodsThe ASID dataset comprises of 397 acute ischemic stroke CT scan images and the Kaggle dataset consists of 200 CT scan images, out of which, 100 images were of the abnormal brain type (hemorrhage), and the remaining 100 images which dont show any abnormalities were categorized as normal cases. The results of LMGP were compared with eight other feature extraction methods using linear, sigmoid, polynomial, and RBF kernels of the SVM classifier. ResultsThe effectiveness of our methodology was evaluated using recall, logarithmic loss, accuracy (ACC), confusion matrix and area under curve (ROC) metrics. The LMGP performed best when RBF-SVM was used for the classification and gave an accuracy of 93% and 95% in the case of five-fold and ten-fold cross-validation respectively. The ten-fold cross-validation gave a precision score of 0.9410, a recall score of 0.9521, an F1 score of 0.9458, and a negative log loss of -0.1645. ConclusionLMGP combines a unique and robust approach of CT scan image feature extraction by combining local information along with gradient change in pixels of the image. The present results of the study suggest the improved performance of the LMGP method over other methods compared in this study effectively.