Brain stroke classification and segmentation using encoder-decoder based deep convolutional neural networks

Accurate diagnosis of brain stroke, classification and segmentation of the stroke are extremely important for physicians to focus on specific points of the brain and apply the right treatment to patients. Encoder-decoder deep learning-based methods have been effectively integrated into many artificial intelligence applications. On the other hand, such networks have many disadvantages due to sampling methods, learning methodologies, and efficient operations. In this study, U-Net, one of the encoder-decoder deep learning-based convolutional neural networks (CNNs), has been developed and proposed for the classification and segmentation of brain stroke. A convolutional deep network architecture is proposed with an optimized dimensional U-Net (D-UNet) by blocking and adaptively sequencing the convolution layers and then optimizing the number of activation functions and hyperparameters. The proposed method examines the computed tomography (CT) images from the dataset used to determine whether there is a brain stroke. It can determine if a stroke is caused by ischemia or hemorrhage once it has occurred. Additionally, the proposed method can precisely reveal the region overlaid by the radiologist and segment the existing stroke. The proposed method is compared with other existing CNN-type architectures by performing various experiments on the same real dataset via Python scripts. The results show that the proposed model performs well, with accuracy rates for stroke classification of 98.9% and ischemia and hemorrhage classification of 98.5%, respectively. Moreover, the segmentation of brain strokes using the proposed model yielded an intersection over union (IoU) rate of 95.2%.