Spatial Distribution-Based Pseudo Labeling for Pathological Image Segmentation

Yuki Shigeyasu, Kengo Araki, Shota Harada, Akihiko Yoshizawa, Kazuhiro Terada, Ryoma Bise

Abstract: This paper proposes a semi-supervised pathological image segmentation method that can improve the segmentation using a large amount of unlabeled data. Typical pseudo labeling methods select pseudo labels from unlabeled data to be used for re-training based on the confidence of each patch. However, the initial estimation is not accurate, so the pseudo labels contain many inaccurate labels. This may affect the segmentation performance. The highlight of this study is that we avoid selecting incorrect pseudo labels by introducing a spatial distribution model in a whole slide image. This is based on the assumption that a tumor region forms a cluster in tissue. This improves pseudo label selection and segmentation performance. Experimental results demonstrate the effectiveness of our method, where our method achieved the best segmentation performance in comparison.