Blood Vessel Segmentation From Low-Contrast and Wide-Field Optical Microscopic Images of Cranial Window by Attention-Gate-Based Network

Yunheng Wu, Masahiro Oda, Yuichiro Hayashi, Takanori Takebe, Shogo Nagata, Cheng Wang, Kensaku Mori; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2022, pp. 1864-1873

The stereomicroscope, which is an optical microscope, is used to observe the organoids cultured in cranial windows. A cranial window is a light accessible observation window made on the brain of mice through craniotomy. Organoids research is often conducted on cranial windows. Hence, the observation of blood vessels in them is important for organoid research, like organoid vascularization. Therefore, achieving a simple, low-cost method that extracts blood vessel structures would significantly help researchers observe the blood vessels in cranial windows from microscopic images. However, wide-field optical microscopic images taken by stereomicroscope suffer from low contrast and dura mater occlusion, complicating the observation of the blood vessels in such images. To address such problems and assist researchers who are observing vascular structures, we propose a method that segments blood vessels in cranial windows from low-contrast and wide-field microscopic images. Our method is based on the Attention U-Net framework and clDice, which considers the connectivity of blood vessels. In addition, for low-contrast and partial occlusion problems, we used contrast enhancement and dehazing as preprocessing steps. Our method achieved a Dice score of 75.56%, a clDice score of 79.95%, and the Accuracy of 91.41% on our microscopic image dataset, suggesting that our method can extract blood vessels from low-contrast and wide-field microscopic images better than other methods.