부산대학교 도서관

We qualitatively compare three image segmentation techniques (filter and threshold-based segmentation, texture-based clustering and deep learning) for histologic images of intracranial aneurysms. Due to remodeling of the vessel wall and aneurysm formation, the tissue is highly diverse. Only the deep learning segmentation provided semantic information about the segmented tissue. The other segmentation techniques were designed to segment areas of different textures and tissues, respectively. Therefore, in contrast to the deep learning approach, they did not require knowledge of all tissue types possible occurring in intracranial aneurysms. Rare tissue classes were missed by the deep learning segmentation, but the resolution of the deep learning segmentation was better than the ground truth segmentation. Overall, the deep learning segmentation of ten classes achieved a test accuracy of 60.68%.