Improvement of Deep-Learning Algorithms for Disease Detection: The Case of Cerebral Hemorrhage

Abstract

Abstract. Cerebral haemorrhage is a serious condition and a major public health issue that requires immediate and accurate care to guide doctors in their treatment decisions. This study developed three deep learning models to accurately identify and classify images of haemorrhages based on normal brain images from computed tomography (CT) scans. These models include two pre-trained models (VGG16 and VGG19) and a custom convolutional neural network (CNN). Due to the severe effects of this disease (paralysis, disability, long-term death) and the challenge of identifying and interpreting it for healthcare professionals, the research considered using these models with a dataset comprising two classes: haemorrhagic and normal. The three models were tested under the same conditions, and the results demonstrated each model's ability to generate data. VGG19 showed 99.8% accuracy, 3% loss, 99% detection and classification capability, and 99% sensitivity. The pre-trained VGG16 model generated an accuracy of 99.7%, an estimated error margin of 30%, a detection capacity rated at 99% and a sensitivity of 98%. The custom CNN model performed the worst, with an accuracy of 89%, an error rate of 88%, a recall of 91% and a lower sensitivity level estimated at 84%. The VGG19 approach performed better than the other models.