Related Experiment Video
Updated: Oct 8, 2025
09:16
Supervised Machine Learning for Semi-Quantification of Extracellular DNA in Glomerulonephritis
Published on: June 18, 2020
7.0K
Deep learning segmentation of glomeruli on kidney donor frozen sections
Xiang Li1, Richard C Davis2, Yuemei Xu2,3
1Duke University, Department of Electrical and Computer Engineering, Durham, North Carolina, United States.
Journal of Medical Imaging (Bellingham, Wash.)
|December 24, 2021
Summary
Deep learning models can automatically quantify nonsclerotic and sclerotic glomeruli in donor kidney biopsies, offering a robust tool for pathologists. This approach aids in evaluating kidney health and improving diagnostic accuracy.
Area of Science:
- Nephrology
- Computational Pathology
- Medical Imaging Analysis
Background:
- Accurate quantification of glomeruli in kidney biopsies is crucial for diagnosis and prognosis.
- Traditional methods are labor-intensive and may lack consistency.
- Advances in computational image analysis present opportunities for automated histologic parameter quantification.
Purpose of the Study:
- To develop and validate deep learning (DL) models for automatic quantification of nonsclerotic and sclerotic glomeruli.
- To assess the performance of DL models in segmenting glomeruli on frozen sections of donor kidney biopsies.
- To compare DL-derived glomerular counts with manual segmentation and standard-of-care counts.
Main Methods:
- Utilized 258 whole slide images (WSIs) from donor kidney biopsies, including institutional and external datasets.
- Manually annotated nonsclerotic and sclerotic glomeruli on all WSIs.
- Developed a nine-layer convolutional neural network (U-Net architecture) for glomeruli segmentation.
- Compared DL-derived segmentation with manual segmentation and reported glomerular counts.
Main Results:
- Achieved average Dice similarity coefficients of 0.90 for nonsclerotic and 0.83 for sclerotic glomeruli.
- Demonstrated high recall and precision scores for both nonsclerotic (0.93, 0.96, 0.90) and sclerotic (0.87, 0.93, 0.81) glomeruli.
- Found DL-derived and manual segmentation counts to be comparable, but statistically different from reported counts.
Conclusions:
- Deep learning segmentation is a feasible and robust method for automatic glomeruli quantification in kidney biopsies.
- This technology represents a significant step towards developing new protocols for evaluating donor kidney biopsies.
- Automated quantification can potentially enhance efficiency and consistency in histopathologic assessments.
