Deep Learning for Intracranial Hemorrhage Detection and Classification in Brain CT Scans: A Systematic Review and Hybrid Model Approach

Abstract

ABSTRACT Intracranial hemorrhage (ICH) is a medical emergency that requires rapid and accurate diagnosis, with non-contrast computed tomography (CT) serving as the primary imaging modality for detecting acute bleeding. In recent years, machine learning and deep learning approaches have been increasingly explored to support automated detection and classification of ICH subtypes, including epidural, subdural, intraparenchymal, intraventricular, and subarachnoid hemorrhages. This systematic review consolidates contemporary machine learning and deep learning methodologies applied to intracranial hemorrhage (ICH) detection and classification from non-contrast CT scans. The reviewed approaches include conventional convolutional neural networks, three-dimensional CNNs, hybrid and ensemble frameworks, and emerging transformer-based architectures. Common preprocessing strategies such as Hounsfield Unit windowing, skull stripping, and data augmentation are examined, alongside explainable AI techniques including Grad-CAM to enhance clinical interpretability. Key methodological trends, quantitative performance outcomes, limitations, and future research directions are discussed to facilitate the translation of automated ICH detection systems into real-world clinical workflows. This systematic review consolidates contemporary machine learning and deep learning methodologies applied to ICH analysis, covering conventional convolutional neural networks (CNNs), three-dimensional CNNs, hybrid architectures, ensemble models, and emerging transformer-based approaches. We also examine commonly employed preprocessing strategies such as Hounsfield Unit–based windowing, skull stripping, and data augmentation, along with explainable AI techniques including Grad-CAM that enhance clinical interpretability. In addition, drawing upon insights from the reviewed literature, we present an illustrative hybrid DenseNet121–LSTM architecture to demonstrate how spatial and sequential feature learning strategies identified in prior studies can be practically integrated within a unified framework. Key challenges, limitations, and future research directions are discussed to facilitate the translation of automated ICH detection systems into real-world clinical workflows.