Pathological image classification via embedded fusion mutual learning. (January 2023)

• Record Type: Journal Article
• Authors: Li, Guangli; Wu, Guangting; Xu, Guangxin; Li, Chuanxiu; Zhu, Zhiliang; Ye, Yiyuan; Zhang, Hongbin
• Is Part Of: Biomedical signal processing and control. Volume 79(2023)Part 2
• Publication Date: 2023-01
• Subjects: Pathological image classification -- Deep mutual learning -- Adaptive feature fusion -- Ensemble classifier; 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.2022.104181
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: EFML is proposed for the classification of the top three cancers. A new adaptive feature fusion is proposed for fusing diverse pathological knowledge. EFML can use heterogeneous or homogeneous networks to generate several variants. EFML handles high-resolution pathological images well. EFML is an efficient and magnification scale independent model. Abstract: Deep learning models have been widely used in pathological image classification. However, most researches employ complex but inefficient neural networks to implement this task. And the implicit but complementary pathological knowledge between heterogeneous networks has not been fully explored. To alleviate these problems, we propose a novel method, named embedded fusion mutual learning (EFML). First, online mutual learning is carried out between two parallel heterogeneous networks, which forms two effective feature extractors. Second, an adaptive feature fusion classifier and an ensemble classifier are embedded into EFML to learn the diverse knowledge from feature maps and logits output, simultaneously. Finally, we combine the logits output and the fused feature maps to jointly supervise model training. We conducted experiments on three public datasets: BreaKHis, BACH and LC25000. Extensive experiments demonstrate the effectiveness, robustness, and generalization ability of EFML. Moreover, EFML is efficient and magnification scale independent. It can get evident performance enhancements on the high-resolution pathological images. All these firmly demonstrate its clinical practicality.