Lightweight deep learning framework for intracranial hemorrhage detection in brain CT scans

Intracranial hemorrhage, including clinically significant intracranial hemorrhage conditions, is a life-threatening condition in which rapid and accurate detection through brain computed tomography (CT) scans is crucial for patient survival. Manual interpretation of these scans remains time-consuming and may vary between observers, leading to potential diagnostic delays. This study investigates the application of lightweight and mobile-optimized deep-learning models for the automated detection of intracranial hemorrhage using a curated subset of a publicly available brain CT hemorrhage dataset, for automated intracranial hemorrhage detection. A custom convolutional neural network (CNN) and two transfer-learning architectures MobileNetV2 and EfficientNet-B0 were evaluated in terms of diagnostic accuracy, generalization, and computational efficiency. Among the tested models, MobileNetV2 demonstrated the highest overall performance, achieving an accuracy of 87% and an AUC of 0.94, while the lightweight CNN achieved 79% accuracy. EfficientNet-B0 also showed competitive results but required greater computational resources. The findings demonstrate that lightweight neural architectures can achieve reliable diagnostic performance while remaining suitable for assistive decision-support applications, although further improvement in sensitivity and clinical validation are required. The study highlights that carefully optimized deep-learning systems can support preliminary clinical assessment; however, additional validation and performance refinement are necessary before practical real-world deployment.