MixMicrobleedNet: segmentation of cerebral microbleeds using nnU-Net

Cerebral microbleeds are hypointense, small, and round or ovoid lesions [1, 2]; visible on magnetic resonance imaging (MRI) with gradient echo, T2*, or susceptibility weighted (SWI) imaging[3, 4, 5]. Assessment of cerebral microbleeds is mostly performed by visual inspection, using validated rating scales such as the Microbleed Anatomical Rating Scale (MARS)[6] or Brain Observer MicroBleed Scale (BOMBS)[7]. In the past decade, prior to the rise of deep learning technology in medical image analysis[8], semi-automated tools to assist with cerebral microbleed detection have been developed. These include techniques based on unified segmentation[9], support-vector machines[10], or the radial symmetry transform[11, 12, 13]. In the more recent years, owing to the great advances provided by deep learning techniques, the number of methods for fully automatic microbleed detection has increased considerably[14, 15, 16, 17, 18]. In this work, we explore the use of nnU-Net[19] as a fully automated tool for microbleed segmentation. This self-configuring deep learning-based semantic segmentation method has shown good performance in a number of international biomedical segmentation competitions[20], but has not been applied to the task of cerebral microbleed detection and segmentation.