An Attention-Based ResNet Architecture for Acute Hemorrhage Detection and Classification: Toward a Health 4.0 Digital Twin Study

Abstract

As AI advances, new applications and services are being developed to enhance current services. With the emergence of digital twin technology (DT), an exact digital replica of any natural system or physical object/state can be created. DT incorporates several machines and deep learning-based approaches to efficiently analyze and process the collected data in order to understand the system pattern better. Consequently, Health 4.0 applications can get advantages from these cutting-edge technologies. In this article, we focus on Intracranial haemorrhage (ICH) which is a life-threatening emergency that needs immediate diagnosis and treatment. ICH is caused by the bleeding inside the skull or brain. It is classified in five main subtypes; a) subdural hemorrhage (SH), b) epidural haemorrhage (EH), c) intraparenchymal haemorrhage (IH), d) subarachnoid haemorrhage (SAH) and e) intraventricular hemorrhage (IVH). Radiologists typically examine computed tomography (CT) scan of the patients to determine the ICH and its subtype. However, manual assessment of the CT scan is a complex and time-consuming task. Pre-trained convolutional neural networks (CNNs) are state-of-the-art for the ICH classification. But they suffer from the curse of dimensionality and use redundant and noisy features which causes memory and computation overhead. The problem of imbalanced data poses the challenge for achieving model generalization. This paper proposes a hybrid attention-based ResNet architecture for ICH detection and classification. Principal component analysis (PCA) is used for dimensionality reduction and redundant feature removal whereas xtreme gradient boosting (XGBoost) is used for ICH classfication. The class imbalance problem is resolved using deep convolutional generative adversarial network (DCGAN). The proposed system is evaluated using the dataset assembled during Radiologist Society of North America (RSNA) ICH detection challenge 2019. The findings show that our proposed model outperforms existing state-of-the-art models in terms of accuracy and F1-score. ICH classification achieved accuracies of 99.3%, 97.1%, 96.7%, and 96.1%, for detecting EH, IH, IVH, SAH, and SH subtypes respectively, while F1-score is 96.1%, which is also best when compared with the benchmark studies.

Topics & Concepts