Deep Learning Technique for Detecting and Analysing Ischemic Stroke using MRI Images

The quantitative examination of images taken from cerebrum MRI assumes a significant part in the process of diagnosis and medication for the strokes. The deep learning and convolutional neural networks with learning capacity supply a powerful mechanism for detection of lesions. To concentrate on the properties of stroke injuries and complete detection activities automatically, we have gathered from various medical means, the data of MRI images of brain, which are belong to different patients affected by ischemic strokes. The classifications of deep learning-based networks used for detecting the objects such as SSD, RCNN-ResNet101, RCNN-VGG16 and YOLOV3 are applied to carry out programmed detection of lesions by achieving the best accuracy compared to existing schemes on Diffused Weight, Flair and T1 modalities of MRI datasets. The technique is developed to extract deep features during encoding stage and features are minimized using layers connected fully. And the significant features which are handcrafted including LBP and GLCM are joined along with the deep features. In order to increase the dimension of feature vector to maximum, the concatenation of feature is implemented by combining deep features and handcrafted features. And this vector is then used to train and test performance of classifiers. In order to classify various modalities of brain images in to normal or stroked, the binary classification is applied in proposed work. In the beginning, SoftMax is used as default for performing the classification. In continuation, the classifiers including SVM, KNN, DT, and RF are utilized. The KNN records accuracy of 97.60% and SVM records accuracy of 98.60%. Depending on selected classifier, performance of the classifiers is evaluated individually and the best outcome is taken in to consideration for confirming the performance of the technique.