Abstract: FR-PO989
A Deep Learning-Based Approach for Glomeruli Object Extraction from Multistained Renal Biopsy Pathologic Images
Session Information
- Pathology and Lab Medicine: Basic
November 08, 2019 | Location: Exhibit Hall, Walter E. Washington Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Pathology and Lab Medicine
- 1601 Pathology and Lab Medicine: Basic
Authors
- Jiang, Lei, Peking University People's Hospital, Beijing, China
- Mei, Ke, Beijing University of Posts and Telecommunications, Beijing, China
- Zhu, Chuang, Beijing University of Posts and Telecommunications, Beijing, China
- Liu, Jun, Beijing University of Posts and Telecommunications, Beijing, China
- Yan, Yu, Peking university people's hospital, Beijing, China
- Dong, Bao, Peking university people's hospital, Beijing, China
- Wang, Gongwei, Peking University People’s Hospital, Beijing, China
- Shi, Hongxia, Peking University People's Hospital, Beijing, China
Background
Glomeruli extraction from pathologic images is a key step in automatic analysis of renal biopsy. We present a deep learning-based approach for the object extraction of three types of glomeruli with various pathological lesions in multi-stained images.
Methods
Sources of images: 1.1947 glomeruli from images captured at 10x, 20x, and 40x, including 33 pathological types of kidney diseases; 2. 601 glomeruli from 44 whole slide images (WSI) scanned at 40x using Precice 500B scanner (UNIC Technologies Inc, China). Slides were stained with Periodic acid-Schiff (PAS), Periodic Acid-silver Methenamine (PASM), and Masson’s trichrome stains. Glomeruli were divided into training and testing sets.
Mask R-CNN architecture based on convolutional neural networks (CNN) was trained by using glomeruli training set. The adopted mask R-CNN, which is built by extending Faster R-CNN by adding a branch for predicting an object mask, can detect, classify, and segment three types of glomeruli: 1. glomerulus with basically normal structure (gn), 2.global sclerosis (gs), and 3.glomerulus with other abnormal structure (gl) at the same time.
Results
The detection and pixel level segmentation results was graded via average precision , average recall, and F-score (true positives were defined by >50% overlap of the predicted region).
Conclusion
We present a robust network using relatively limited sample size, which can detect normal and abnormal glomeruli stained with PAS, Masson and PASM.
| Type | Average precision | Average recall | F-score |
| gn | 0.903 | 0.862 | 0.882 |
| gs | 0.746 | 0.676 | 0.709 |
| gl | 0.750 | 0.689 | 0.718 |
| All | 0.796 | 0.742 | 0.768 |
| Detection of total glomeruli | 0.898 | 0.952 | 0.924 |
Funding
- Clinical Revenue Support