Abstract: TH-PO088
Identification of Glomeruli for Improved Enumeration in Renal Biopsies Using Convolutional Neural Networks and Deep Learning
Session Information
- Clinical/Diagnostic Renal Pathology and Lab Medicine - I
November 02, 2017 | Location: Hall H, Morial Convention Center
Abstract Time: 10:00 AM - 10:00 AM
Category: Glomerular
- 1004 Clinical/Diagnostic Renal Pathology and Lab Medicine
Authors
- Street, Jonathan, National Institute of Diabetes and Digestive and Kidney Diseases, Bethesda, Maryland, United States
- Bellomo, Tiffany Rose, National Institute of Diabetes and Digestive and Kidney Diseases, Bethesda, Maryland, United States
- Koritzinsky, Erik H., National Institute of Diabetes and Digestive and Kidney Diseases, Bethesda, Maryland, United States
- Hewitt, Stephen M., National Cancer Institute, Bethesda, Alabama, United States
- Yuen, Peter S.T., National Institute of Diabetes and Digestive and Kidney Diseases, Bethesda, Maryland, United States
- Star, Robert A., National Institute of Diabetes and Digestive and Kidney Diseases, Bethesda, Maryland, United States
Background
Accurate detection and counting of glomeruli in renal biopsies is important to assess biopsy adequacy, and for diagnostic accuracy (for example, the percentage of sclerotic glomeruli). A recent evaluation indicates that traditional methods can under-count glomeruli by ~50% when compared to more labor-intensive annotation methods. In recent years, deep learning has significantly advanced in a variety of tasks including image recognition, classification, and segmentation. To improve the accuracy of glomerular enumeration in renal biopsy evaluation, we are developing a convolutional neural network/deep learning model to locate glomeruli.
Methods
Twenty-two biopsy sections stained with hematoxylin and eosin were imaged using whole slide scanners (Aperio and Hamamatsu) for use in training and validation of the model. A custom application was developed to record the position of glomeruli in each biopsy with masks processed as geometries (boundaries stored rather than each pixel). This approach minimized file sizes and supported complex operations. Training and validation was conducted at the NIH high-performance computing facility employing multiple NVidia K20 GPU equipped nodes. The keras software package was used to define and train the neural network. Data augmentation and a modestly sized network (3 convolutional layers, a single hidden fully connected layer and an output layer) were used to optimize performance without overfitting the dataset.
Results
The accuracy of the neural network model was 91%. The model performance was robust to changes in stain intensity and the scanner used for imaging. Importantly, manual review of disagreements revealed errors in annotation (the model correctly identified glomeruli that had been missed by the human annotator).
Conclusion
Deep learning techniques can be utilized to accurately identify glomeruli in renal biopsies and can be scaled to handle the large images generated by whole slide scans. A convolutional neural network may help to improve the accuracy of glomeruli localization and enumeration, without the need for human intervention. This model could be leveraged for classification of glomerular morphology.
Funding
- NIDDK Support