Histopathological image classification based on cross-domain deep transferred feature fusion. (July 2021)

• Record Type: Journal Article
• Authors: Wang, Pin; Li, Pufei; Li, Yongming; Wang, Jiaxin; Xu, Jin
• Is Part Of: Biomedical signal processing and control. Volume 68(2021)
• Publication Date: 2021-07
• Subjects: Breast cancer histopathological images -- Cross-domain transfer learning -- Feature fusion; 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.2021.102705
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: A novel deep network with residual learning is designed to extract multiple-layer features simultaneously. A cross-domain deep feature fusion method is proposed to minimize the discrepancy between the source and the target domain, and fuse different hierarchical layer features to obtain more discriminative features. A regularization mechanism is adopted to adaptively optimize the fusion weights. Abstract: Transfer learning-based methods for breast cancer histopathological images classification are widely used for computer-aided cancer diagnosis. However, most existing methods directly migrate the model pre-trained on natural images to medical images with little consideration of the differences in the distribution of source and target domain data features. And combining multi-layer features can obtain more discriminative features which is helpful for improving the classification performance. In this paper, a novel histopathological images classification method based on cross-domain transfer learning and multi-stage feature fusion (Cd-dtffNet) is proposed. First, a novel network with residual learning is designed which can extract features from multiple levels. Then, the ability of the network to extract local features is learned by migrating from the source domain to the bridge domain, and the ability to extract global features is learned by migrating from the bridge domain to the target domain. Moreover, a feature fusion strategy with the L2 regularization term is utilized to fuse the extracted local and global features. The proposed method Cd-dtffNet was tested on the breast cancer histopathological image dataset. Experimental results (normal vs malignant: 99.09 %, normal vs uninvolved: 97.71 %, normal vs malignant + uninvolved: 98.27 %) demonstrate effectiveness of the proposed method for breast cancer classification in clinical diagnosis.