Deep learning models are widely used for solving problems in different applications. Especially Convolutional Neural Network (CNN) based models are found suitable for medical image analysis. As brain stroke is increasing in alarming rate, it is essential to have better approaches to detect it in time. Brain MRI is one of the medical imaging technologies widely used for brain imaging.we proposed certain advancements to well-known deep learning models like VGG16, ResNet50 and DenseNet121 for enhancing brain stroke detection performance. These models are optimized based on the brain stroke detection problem in hand as they are not specialized for a specific problem. We proposed an algorithm, named Deep Efficient Stroke Detection (ESD), that exploids enhanced deep learning models in pipeline. The experimental results revealed that there is performance improvement with optimized models. Highest accuracy is achieved by ResNet50 with 95.67%.

Deep learning has emerged to be efficient Artificial Intelligence (AI) phenomena to solve problems in healthcare industry. Particularly Convolutional Neural Network (CNN) models have attracted researchers due to their efficiency in medical image analysis. According to World Health Organization (WHO), rapidly developing cerebral malfunction, brain stroke, is the second leading cause of death across the globe. Brain MRI scans, when analysed quantitatively, play vital role in diagnosis and treatment of stroke. There are many existing methods built on deep learning for stroke diagnosis. However, an automatic, reliable and faster method that not only helps in stroke diagnosis but also demarcate affected regions as part of Clinical Decision Support System (CDSS) is much desired. Towards this objective, we proposed an Automated Deep Learning based Brain Stroke Detection Framework (ADL-BSDF). It does not rely on expertise of healthcare professional in diagnosis and know the extent of damage enabling physician to make quick decisions. The framework is realized by two algorithms proposed. The first algorithm known as CNN-based Deep Learning for Brain Stroke Detection (CNNDL-BSD) focuses on accurate detection of stroke. The second algorithm, Deep Auto encoder for Stroke Severity Detection (DA-SSD), focuses on revealing extent of damage or severity of the stroke. The framework is evaluated against state of the art deep learning models such as EfficientNet, ResNet50 and VGG16.