Abstract: PO2244
Deep Neural Network Facilitated Immunofluorescence Assessment of Glomerular Diseases: A Preliminary Report
Session Information
- Pathology and Lab Medicine: Clinical
October 22, 2020 | Location: On-Demand
Abstract Time: 10:00 AM - 12:00 PM
Category: Pathology and Lab Medicine
- 1602 Pathology and Lab Medicine: Clinical
Authors
- Xia, Peng, Nephrology Division, Peking Union Medical College Hospital, Beijing, China
- Lv, Zhilong, Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China
- Wen, Yubing, Nephrology Division, Peking Union Medical College Hospital, Beijing, China
- Zhao, Xuesong, Nephrology Division, Peking Union Medical College Hospital, Beijing, China
- Zheng, Hua, Nephrology Division, Peking Union Medical College Hospital, Beijing, China
- Li, Xuemei, Nephrology Division, Peking Union Medical College Hospital, Beijing, China
- Ren, Fei, Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China
- Chen, Limeng, Nephrology Division, Peking Union Medical College Hospital, Beijing, China
Background
Immunofluorescence (IF) tests of renal tissue are important in diagnosing glomerular diseases. Deep neural network were used to facilitate analysis of pathologic images recently, but not in IF assessment. We proposed a novel Convolutional Residual Dense Network (CR-DenseNet) to facilitate IF assessment of glomerular diseases.
Methods
A dataset with 614 IF images of glomerulus, including IgA Nephropathy, IgAN (n=319), Idiopathic Membranous Nephropathy, IMN (n=211) and Secondary Membranous Nephropathy, SMN (n=84) from Peking Union Medical College Hospital, PUMCH were used for training of CR-DenseNet. Additional 78 IF images from PUMCH (35 IgAN, 15 IMN and 28 SMN) and 98 IF images (36 IgAN, 34 IMN and 28 SMN) from other 3 hospitals were used for validation and human tests. These images were annotated by two nephropathologists independently. Convolutional residual dense blocks were introduced. Each of them consisted of a dense block with a convolutional skip connection to fully exploit the dense local features (Figure 1). Performance was evaluated by overall accuracy, sensitivity, specificity and F1 score. F1 was computed as 2×True Positive, TP / (2×TP + False Positive + False Negative).
Results
The proposed CR-DenseNet outperformed the state-of-the-art method with overall 85.3% recognition accuracy. For the validation data from PUMCH, the sensitivity of recognizing IgAN, IMN and SMN were 78.3% to 93.3% and the specificity were 88.9% to 97.4%. For the validation dataset from other hospitals, the sensitivity and specificity were 79.2% to 93.5% and 88.5% to 98.2%, respectively. Eight nephrologists' average accuracy for recognizing each classification were 50.0% to 76.0%. The F1 score were 0.542 to 0.802. CR-DenseNet demonstrated superior performance. The corresponding accuracy and F1 score were 80.0% to 84.6% and 0.762 to 0.880, respectively.
Conclusion
Our data showed that CR-DenseNet model were useful in IF assessment of typical glomerular diseases.
CR-DenseNet Overview
Funding
- Government Support - Non-U.S.