Diagnosis of endometrium hyperplasia and screening of endometrial intraepithelial neoplasia in histopathological images using a global-to-local multi-scale convolutional neural network

• Published in: Computer methods and programs in biomedicine 2022-06, Vol.221, p.106906-106906, Article 106906
• Main Authors: Zhao, Fengjun, Dong, Didi, Du, Hongyan, Guo, Yinan, Su, Xue, Wang, Zhiwei, Xie, Xiaoyang, Wang, Mingjuan, Zhang, Haiyan, Cao, Xin, He, Xiaowei
• Format: Article
• Language: English
• Subjects: Computer-aided-diagnosis; Convolutional neural network; Endometrial hyperplasia; Endometrial intraepithelial neoplasia; Histopathological image
• ISSN: 0169-2607; 1872-7565
• DOI: 10.1016/j.cmpb.2022.106906

•Diagnosing different endometrial hyperplasia is crucial for the health of females.•Computer-aided-diagnosis with deep learning shows potential for lesion diagnosis.•A multi-scale network combines global and local features from histological images.•Features are extracted from two parallel networks and fused by attention modules.•Proposed model is promising in identifying endometrial intraepithelial neoplasia. Endometrial hyperplasia (EH), a uterine pathology characterized by an increased gland-to-stroma ratio compared to normal endometrium (NE), may precede the development of endometrial cancer (EC). Particularly, atypical EH also known as endometrial intraepithelial neoplasia (EIN), has been proven to be a precursor of EC. Thus, diagnosing different EH (EIN, hyperplasia without atypia (HwA) and NE) and screening EIN from non-EIN are crucial for the health of female reproductive system. Computer-aided-diagnosis (CAD) was used to diagnose endometrial histological images based on machine learning and deep learning. However, these studies perform single-scale image analysis and thus can only characterize partial endometrial features. Empirically, both global (cytological changes relative to background) and local features (gland-to-stromal ratio and lesion dimension) are helpful in identifying endometrial lesions. We proposed a global-to-local multi-scale convolutional neural network (G2LNet) to diagnose different EH and to screen EIN in endometrial histological images stained by hematoxylin and eosin (H&E). The G2LNet first used a supervised model in the global part to extract contextual features of endometrial lesions, and simultaneously deployed multi-instance learning in the local part to obtain textural features from multiple image patches. The contextual and textural features were used together to diagnose different endometrial lesions after fusion by a convolutional block attention module. In addition, we visualized the salient regions on both the global image and local images to investigate the interpretability of the model in endometrial diagnosis. In the five-fold cross validation on 7812 H&E images from 467 endometrial specimens, G2LNet achieved an accuracy of 97.01% for EH diagnosis and an area-under-the-curve (AUC) of 0.9902 for EIN screening, significantly higher than state-of-the-arts. In external validation on 1631 H&E images from 135 specimens, G2LNet achieved an accuracy of 95.34% for EH diagnosis, which was comparable to that of a mid-level