Classification of Hemorrhagic Stroke Lesions Based on CT Images and Machine Learning Algorithms. A Study on a Highly Imbalanced Dataset

Hemorrhagic stroke is a common disease that has a high mortality rate. The patient’s chances of a complete recovery depend on where in the brain the hemorrhage occurred. Automatic classification based on machine learning (ML) techniques is often proposed in clinical applications, depending on the availability of data and computational resources. The traditional ML algorithms, such as Random Forest and Support Vector Machine, are often applied to tabular data requiring smaller datasets and less intensive computational power, while the more advanced, but costly techniques, such as convolutional neural networks, are frequently applied to image data. This paper describes the automatic classification of the 5 subtypes of hemorrhagic stroke, namely Intraparenchymal (IPH), Subarachnoid (SAH), Intraventricular (IVH), Epidural (EDH) and Subdural (SDH), comparing the results after applying traditional machine learning and deep learning (DL) methods, on the same dataset. The detection of the 5 classes had 74% balanced accuracy for the best identified ML algorithm and 90% after applying a DL method.