Classification for Intracranial Hemorrhage Based on Deep Learning

Abstract:
Intracranial hemorrhage has become a social health problem that cannot be ignored, which leads to huge strokes worldwide and serious sequelae. As the first choice to diagnose the Intracranial hemorrhage, attention to subtle details on CT images and speed of diagnosis are key points in medical diagnosis. In this paper, a hemorrhage detection model based on convolutional neural network (CNN) was established to realize high accuracy classification based on specific CT images dataset which were classified into two categories with label hemorrhage and no hemorrhage. The original model had 3 convolutional layers and 2 fully-connected layer. To improve the CNN model, several controlled experiments were conducted to figure out the influence on the testing accuracy with different number of convolutional layers, different number of neurons in fully-connected layers and different number of fully-connected layers. The experiments are carried out in the order mentioned above, and the basic model of the subsequent optimization experiments is built on the best performing model in the previous experiments. To verify the performance of self-designed CNN model, MobileNetV2 and VGG16 were selected and trained on the same dataset. The results indicated that self-designed CNN model had better performance than that of MobileNetV2 and VGG16 on collected dataset.