Automatic Hemorrhage Detection In Magnetic Resonance Imaging In Cerebral Amyloid Angiopathy

Cerebral hemorrhages, or intracranial hemorrhages, can be caused by the rupture of blood vessels inside the skull. One of the most common causes of brain hemorrhage is cerebral amyloid angiopathy, which can also be associated with Alzheimer's disease. However, the Magnetic Resonance Imaging (MRI) scans in which the hemorrhages can be found must be manually examined and segmented which is very challenging due to the difficult differentiation between hemorrhages and iron deposits or calcifications which makes this a tiresome process susceptible to human error. To improve and automate the detection of brain hemorrhages, deep artificial neural networks were used as they have shown good results in similar applications. The dataset used in this work contains MRI data from 65 patients, using only T2* scans. We propose a two-stage approach in which manual annotation of T2* examinations were used to train a network to identify exams that contain a specific type of hemorrhage, either micro- or macro-hemorrhages. The developed approach achieved an accuracy of 0.94 and 0.9 in detecting slices containing macro and micro hemorrhage, respectively, by using a DenseNet architecture. In the second stage, the identified scans are segmented using an AttentionUnet network which achieved a Dice Score of 0.77 for macro hemorrhages and 0.66 for micro was achieved. This solution provides good results, which proved sufficient for individual detection of hemorrhages however, it still has room for further improvement.