Abstract: FR-PO974

Automatic Detection of Global Sclerosis in Pathological Glomerular Images with Deep Neural Networks

Session Information

Category: Bioengineering and Informatics

  • 101 Bioengineering and Informatics


  • Uchino, Eiichiro, Kyoto University Graduate School of Medicine, Kyoto, Japan
  • Taneishi, Kei, RIKEN Advanced Institute for Computational Science, Kobe, Japan
  • Sato, Noriaki, Kyoto University Graduate School of Medicine, Kyoto, Japan
  • Yokoi, Hideki, Kyoto University Graduate School of Medicine, Kyoto, Japan
  • Okuno, Yasushi, Kyoto University Graduate School of Medicine, Kyoto, Japan
  • Yanagita, Motoko, Kyoto University Graduate School of Medicine, Kyoto, Japan

Renal pathology is essential for diagnosis and treatment planning for patients with renal disease, and the application of digital pathology, which is a recent active field of automated histology analysis, to renal field is an important challenge. In recent literature, several studies have developed algorithm for detecting glomeruli from a whole slide image of kidney, but few studies have focused on detecting pathological features such as global sclerosis from those glomerular images.


We developed deep convolutional neural network (CNN) models which can automatically classify glomerular images as global sclerosis or not. Total 12422 glomerular images of periodic acid-Schiff (PAS) stain (n=3180), periodic acid methenamine silver (PAM) stain (n=3311), hematoxylin and eosin (HE) stain (n=2979), and Masson's trichrome (MT) or elastica-Masson (EM) stain (n=2952) were cropped from whole slide images of 133 patients who underwent renal biopsy in our hospital in 2013 and 2014, and were manually annotated based on clinical pathology reports. By each staining, CNNs were trained and validated by 5-fold cross validation with five patient groups assigned randomly. We evaluated area under the curve (AUC) of receiver operating characteristic (ROC), sensitivity, and specificity at optimal cut-point estimated from Youden's index.


Out of 12422 glomeruli, 1743 (14.0%) were manually annotated as global sclerosis. For detecting global sclerosis, our CNN models for PAS stain had the best AUC in all staining (AUC, 0.977 ± 0.006 (mean ± SD); sensitivity, 95.8 ± 2.6%; specificity, 92.1 ± 2.3%). The models for the other staining showed similar performances for PAM (AUC, 0.971 ± 0.010; sensitivity, 90.3 ± 2.5%; specificity, 95.9 ± 2.4%), HE (AUC, 0.965 ± 0.010; sensitivity, 87.6 ± 2.3%; specificity, 95.4 ± 1.8%), and, MT or EM (AUC, 0.976 ± 0.010; sensitivity, 91.2 ± 3.2%; specificity, 95.2 ± 4.1%).


Our CNN models achieved good performance in detecting global sclerosis from glomerular images, which suggests that this approach is applicable for detecting other pathological features in future studies.