Enhancing Cerebral Microbleed Segmentation with Pretrained UNETR++

Accurate segmentation of cerebral microbleeds (CMBs) is important for diagnosing small vessel diseases, yet it presents significant challenges due to their tiny size and especially the high risk of false positives (e.g., calcifications and blood flow in pial vessels) in magnetic resonance imaging (MRI). While existing studies have employed multi-stage deep neural networks to reduce false positives, these approaches often rely heavily on the performance of the false positive reduction module. To address this and move towards an end-to-end learning scheme, we propose a novel approach that incorporates self-supervised learning through masked image modeling to enrich both encoder and decoder features in UNETR++ using gradient recalled echo T2*-weighted MRI. Noting that existing pre-training strategies often focus on the encoder while neglecting the decoder, our approach pre-trains both components by introducing segmentation and reconstruction heads. We evaluated our model on an in-house dataset and external validation set, demonstrating superior performance compared to the state-of-the-art nnUNet and nnDetection. Additionally, ablation studies revealed that pre-training both the encoder and decoder subsequently benefits the overall performance of our framework. Our code is available at https://github.com/junmokwon/UNETRppCMBSeg.