COCTA: A Cortex Vessel Segmentation Benchmark Using Optical Coherence Tomography Angiography

Jiaxiang Ren, Wensheng Cheng, Qiaochu Wang, Zhenghong Li, Rutwik Segireddy, Raghuveer Tatineni, Hyomin Jeong, Yanzuo Liu, Ruoyu Xu, Mira Xu, Xiaoyu Hu, Keying Jia, Congwu Du, Yingtian Pan, Haibin Ling

Abstract: Cortical vasculature plays an important role in neuroscience, and it is popularly analyzed using optical coherence tomography angiography (OCTA). Despite two decades of development, however, there remains a lack of high quality vessel segmentation benchmark for cortex OCTA data. In this paper, we introduce a novel OCTA vessel segmentation benchmark of mouse cortex, called COCTA, which largely facilitates the training and evaluation of segmentation models for brain vasculature possible. OCTA involves different types of real-world noise, including speckle noise, motion artifacts and background noise. These heterogeneous noise sources exhibit diverse characteristics and post challenges for current segmentation methods. Therefore, models trained on public retina datasets can not easily generalize to cortex OCTA. Additionally, it is hard to accurately delineate vessel boundaries using mouse so we introduce the stylus pen to ensure a refined and natural-look mask. Besides, our annotators make extensive efforts to remove artifacts and reveal underling vessels. With these corrected manual masks, our dataset is suitable for evaluation in the denoising community. Lastly, we benchmark and analyze various segmentation methods, including convolutional neural networks and transformers, providing insights for the development of new approaches in OCTA. The dataset with manual annotations are available at https://github.com/reckdk/COCTA-dataset.