Cerebral microbleeds are small perivascular haemorrhages in the brain tissue. Microbleeds are one of the markers of presence of cerebrovascular pathology, and are associated with increased risk of dementia and cognitive decline. Currently, they are manually labeled by expert radiologists. However, this task is laborious, time-consuming, and subject to inter and intra rater variability.
In this project, we developed an automated tool to segment microbleeds from T1-weighted, T2-weighted, and T2* images. We used manual segmentations from 78 participants to train a ResNet50 network to detect microbleeds. Due to the small size of our microbleed dataset, we first trained the network to learn another MR image segmentations task (i.e. cerberospinal fluid versus background segmentation), and then used transfer learning to train the network for our task of interest. We assessed the impact of patch size, freezing weights of the initial layers, mini-batch size, and learning rate on the performance of the ResNet50 network.