病理組織画像における細胞膜の閉領域抽出

Nuclear and cell areas are necessary to calculate the N/C ratio which is a useful information for histological diagnosis
especially for cancer. Of the nuclear and cell areas, the cell areas can be calculated by segmenting the cell membranes as closed
areas. Conventional cell membrane segmentation networks, however, have a problem that the accuracy for segmenting cell
membranes as closed areas is low. In this paper, we propose a method to recursively apply Recursive additive complement
network (RacNet) to conventional cell membrane segmentation results. The RacNet is composed of a complement network and
a process in which completion is performed additively only within a certain region in the direction of the cell membrane
extension. Unstained hepatic sections were used in the experiment. The accuracy for segmenting cell membranes as closed areas
was improved up to 89.9% using the RacNet from 50.7% using the cell membrane segmentation network, showing the
effectiveness of the RacNet. The usefulness of bright-field, dark-field, and phase-contrast imaging, which are readily available
in the optical microscope used by pathologists, was compared. Among the three imaging methods, phase-contrast imaging was
the most useful. The highest accuracy was obtained when both dark-field and phase-contrast images were used.