Enhancing point annotations with superpixel and confident learning guided for improving semi-supervised OCT fluid segmentation

Tengjin Weng, Yang Shen, Kai Jin, Yaqi Wang, Zhiming Cheng, Yunxiang Li, Gewen Zhang, Shuai Wang

Abstract: Highlights•We propose a semi-supervised SCLGPA-Net, constructed by a teacher-student architecture. The proposed method can significantly boost model performance while reducing the need for precise annotations.•We propose SGPLG, a superpixel-guided method for generating pseudo-labels and label trust maps based on point annotation. The label trust maps constrain the network from fitting label errors by assigning lower confidence to suspected noisy label pixels.•To our best knowledge, this is the first research that applies the label-denoising method to OCT fluid segmentation.