Abstract: FR-PO990
Deep Learning-Based Segmentation of Normal Histologic Kidney Primitives on Whole Slide Images from NEPTUNE Digital Renal Biopsies
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: Glomerular Diseases
- 1203 Glomerular Diseases: Clinical, Outcomes, and Trials
Authors
- Jayapandian, Catherine P., Case Western Reserve University, Cleveland, Ohio, United States
- Chen, Yijiang, Case Western Reserve University, Cleveland, Ohio, United States
- Janowczyk, Andrew, Case Western Reserve University, Cleveland, Ohio, United States
- Palmer, Matthew, University of Pennsylvania, Philadelphia, Pennsylvania, United States
- Zee, Jarcy, Arbor Research Collaborative for Health, Ann Arbor, Michigan, United States
- Cassol, Clarissa Araujo, The Ohio State University, Columbus, Ohio, United States
- Sekulic, Miroslav, Case Western Reserve University, Cleveland, Ohio, United States
- Hodgin, Jeffrey B., The University of Michigan, Ann Arbor, Michigan, United States
- Hewitt, Stephen M., National Cancer Institute, Bethesda, Maryland, United States
- O'Toole, John F., Cleveland Clinic, Cleveland, Ohio, United States
- Sedor, John R., Cleveland Clinic, Cleveland, Ohio, United States
- Barisoni, Laura, Duke University, Durham, North Carolina, United States
- Madabhushi, Anant, Case Western Reserve University, Cleveland, Ohio, United States
Background
The establishment of digital pathology repositories, such as Nephrotic Syndrome Study Network (NEPTUNE), enables large scale analyses of renal biopsies by sophisticated computational imaging approaches and machine-human interactive protocols. Here we evaluate the performance of U-Net deep learning algorithm for identification of normal histologic primitives in whole slide images (WSIs) across multiple stains.
Methods
Eighteen U-Nets were trained to segment: (i) normal glomerular tufts, (ii) normal glomerular unit (tuft + Bowman’s space and capsule), (iii) normal proximal tubular segments (PT), (iv) normal distal tubular segments (DT), (v) interstitial capillaries, and (vi) arteries. Regions were extracted from 419 WSIs, including 103 H&E, 112 PAS, 100 Silver, 103 Trichrome) from 125 NEPTUNE digital renal biopsies with a diagnosis of Minimal Change Disease. The renal biopsies were randomly sampled into training, validation and testing sets in the ratio 6:1:3. Five pathologists provided the manual segmentation (ground truth). Detection and segmentation results were evaluated using F-Score, True Positive Rate (TPR), Positive Predictive Value (PPV) and Dice Similarity Coefficient (DSC), respectively.
Results
PAS stained WSIs yielded the best performance for all primitives with F-Score (i) 0.93, (ii) 0.94, (iii) 0.91, (iv) 0.93, (v) 0.93 and (vi) 0.85.
Conclusion
This work represents a solid foundation towards enlisting machine learning classifiers to aid large scale tissue quantification efforts. Ongoing effort is devoted to segment abnormal histologic primitives for the development of image-based predictors of disease prognosis.
Funding
- NIDDK Support