Visualization and Interpretation of Nuclear Features in a Deep Learning-Discriminant Analysis of Lobular Endocervical Glandular Hyperplasia with VGG16 and Linear Support Vector Machine.

INTRODUCTION: Lobular endocervical glandular hyperplasia (LEGH), which is considered a precursor lesion of gastric-type adenocarcinoma, exhibits subtle nuclear atypia in cytological specimens and, thus, is occasionally difficult to distinguish from normal endocervical (EC) cells. In the present study, support vector machine (SVM) and deep learning (DL) using the VGG16 model were applied to investigate whether accurate classification is achievable. METHODS: Papanicolaou-stained microscopic images from 56 EC cases and 47 LEGH cases were used. Several features were extracted from each nucleus, and a classification analysis was performed using SVM and DL. Although cross-validation was performed on a case-by-case basis, accuracy was calculated at the image level because a large number of nuclear images were obtained from each case. 　Gradient-weighted Class Activation Mapping (Grad-CAM) was used to visualize attention regions in DL. RESULTS: For the discrimination between EC and LEGH, the accuracies of SVM and DL were 84.2% and 81.2%, respectively. Accuracies between typical LEGH and atypical LEGH were 64.3% for SVM and 54.0% for DL, showing the largest difference. Grad-CAM visualized the differences in attention regions between EC and LEGH. CONCLUSION: The accuracy of SVM, which uses human-selected features, was slightly higher than that of DL, which automatically extracts and uses features without human intervention. This is presumably a result of humans selecting features that are more effective for discrimination. Grad-CAM showed that, in LEGH, the model evaluated the entire nuclear chromatin and focused on chromatin distribution patterns that are difficult for the human eye to discern.