An End-to-End Deep Learning Framework for Predicting Hematoma Expansion in Hemorrhagic Stroke Patients from CT Images

Background and Objectives: Hematoma expansion (HE) occurs in 20% of patients with hemorrhagic stroke within 24 hours of onset and it is associated with a poorer patient outcome. From a clinical point of view, predicting the HE from the initial patient computed tomography (CT) image would be very useful, not only to be able to act faster but to improve therapeutic decisions. Recently, some works have started to study the prediction of the HE as a classification problem to point out if the hematoma will grow or not. However, there are not many works aiming to predict the final extent and appearance of the hematoma.Methods: In this work, we propose an end-to-end deep learning framework for predicting the final hematoma expansion and its corresponding lesion mask. We also explore the problem of having limited data and propose to augment the available dataset with the generation of synthetic images. The proposed approach is evaluated in different scenarios such as training using only the original CT images, using standard data augmentation techniques and using the image synthesis. We perform our experiments in a leave-one-out manner on an in-house dataset consisting of 27 hemorrhagic stroke patients with HE, where longitudinal CT images together with manual lesion annotations were available.Results: The obtained results showed a promising HE prediction, improving the performance when incorporating the use of synthetic images into the model. Our best model obtained a mean Dice score of the HE growth area of 0.506 and an average prediction error of hematoma volume of -3.44 ml.Conclusions The proposed approach showed a great potential for robust prediction of HE, highlighting also the importance of further investigations in this research line that could improve the management of hemorrhagic stroke patients in clinical practice.