Fractal graph convolutional network with MLP-mixer based multi-path feature fusion for classification of histopathological images. (February 2023)

• Record Type: Journal Article
• Authors: Ding, Saisai; Gao, Zhiyang; Wang, Jun; Lu, Minhua; Shi, Jun
• Is Part Of: Expert systems with applications. Volume 212(2023)
• Publication Date: 2023-02
• Subjects: Histopathological images -- Fractal graph convolutional network -- Feature fusion -- MLP-mixer -- Non-local attention; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2022.118793
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: A novel fractal GCN is proposed for classification of histopathological images. FGCN effectively learn multi-level spatial features from histopathological images. An MLP-mixer based multi-path feature fusion unit (MMFFU) is developed. MMFFU effectively fuses multi-level graph representations in FGCN. Abstract: The spatial information among different tissue components and multi-level features is important in histopathological images for pathologists to diagnose cancers. Graph convolutional network (GCN) can effectively learn these spatial features to improve the performance of computer-aided diagnosis (CAD) for histopathological images. The newly proposed GCN-based framework can effectively avoid the complex image preprocessing for graph construction, which integrates convolutional neural network (CNN) and GCN into a framework (named CNN-GCN). However, existing GCN generally cannot well learn multi-level graph features to further promote spatial feature representation. In this work, a new fractal GCN (FGCN) is proposed, which integrates the graph convolution into the fractal structure with different paths to learn multi-level graph representations. Moreover, a novel MLP-mixer-based Multi-path Feature Fusion Unit (MMFFU) is developed in FGCN to fuse these multi-level graph features from multiple paths. In MMFFU, the Mixer layer of MLP-mixer and non-local attention operation are designed to enhance the information communication of features, which further improves feature representation. The proposed FGCN is then embedded into the CNN-GCN framework (named CNN-FGCN) to perform the end-to-end classification of histopathological images. The experimental results on two public histopathological image datasets indicate the effectiveness of CNN-FGCN.