Manifold reconstructed semi-supervised domain adaptation for histopathology images classification. (March 2023)

• Record Type: Journal Article
• Authors: Li, Yongming; Xu, Jin; Wang, Pin; Li, Pufei; Yang, Gongxin; Chen, Rui
• Is Part Of: Biomedical signal processing and control. Volume 81(2023)
• Publication Date: 2023-03
• Subjects: Semi-supervised domain adaptation -- Whole slide image -- Manifold reconstruction -- Embedding features; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.104495
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: A novel pre-trained network BreNet is designed to extract multi-level features. A multi-level feature fusion alignment criterion is proposed to enrich classification beneficial information and measure the discrepancy between source and target domains in the fusion feature space. A novel manifold reconstructed domain adaptation method is proposed to generate the low-dimensional embedding of the multi-level fusion feature and minimize the cross-domain discrepancy in the low-dimensional manifold feature space. Abstract: Deep learning has been applied in computer-aided whole slide images diagnosis widely. However, most deep networks require lots of labeled samples, which is time-consuming and laborious. Meanwhile, most existing methods only extract the deep abstract features only, ignoring the low-level features, which contain the cell structure information. To solve the problems mentioned above, a manifold reconstructed semi-supervised domain adaptation model is proposed for whole slide images' classification. First, a transferred network BreNet is proposed for extracting features from multiple layers of source and target domains. Multi-level features are fused and aligned to characterize the cell structure of the target domain. A novel manifold reconstructed domain adaptation method is proposed to obtain the low-dimensional embedding of the fused features and minimize the cross-domain discrepancy simultaneously. The patch-level prediction probabilities are aggregated for final image-level classification. The experimental results evaluated on three breast cancer datasets indicate the potential of the proposed method for histopathology classification in clinical setting.