ASN's Mission

ASN leads the fight to prevent, treat, and cure kidney diseases throughout the world by educating health professionals and scientists, advancing research and innovation, communicating new knowledge, and advocating for the highest quality care for patients.

learn more

Contact ASN

1401 H St, NW, Ste 900, Washington, DC 20005

email@asn-online.org

202-640-4660

The Latest on Twitter

Kidney Week

Abstract: SA-PO1074

Segmentation Outperforms Classification to Enumerate Glomeruli in Whole Slide Renal Biopsy Images Using Deep Learning

Session Information

Category: Pathology and Lab Medicine

  • 1502 Pathology and Lab Medicine: Clinical

Authors

  • Street, Jonathan, NIDDK/NIH, Bethesda, Maryland, United States
  • Chari, Rohit R., NIDDK/NIH, Bethesda, Maryland, United States
  • Bellomo, Tiffany Rose, NIDDK/NIH, Bethesda, Maryland, United States
  • Hewitt, Stephen M., NCI/NIH, Bethesda, Maryland, United States
  • Yuen, Peter S.T., NIDDK/NIH, Bethesda, Maryland, United States
  • Star, Robert A., NIDDK/NIH, Bethesda, Maryland, United States
Background

Accurate detection and counting of glomeruli in renal biopsies is used to assess biopsy adequacy, and for diagnosis (for example, the percentage of sclerotic glomeruli). Rosenberg et al. recently showed that standard, manual histopathology methods can under-count glomeruli by ~50% compared to more labor-intensive annotation methods. Deep learning has significantly advanced image analysis with previously impractical technologies, such as facial recognition, now possible. We have applied these advances to classification and segmentation of glomeruli. Classification separates small, overlapping image tiles into two classes [glomerulus vs. no glomerulus]. Segmentation separates individual pixels in an image into two classes [inside glomerulus vs outside glomerulus]. Segmentation is similar to masking or finding the border of a glomerulus.

Methods

Fifteen biopsy H&E stained sections were digitized with whole slide scanners (Aperio and Hamamatsu) for training and validation of the model. The outline of each glomerulus was annotated by two people using a custom application we developed. A network with 3 convolutional layers was used for classification and a 7 layer U-Net structure used for the harder problem of segmentation.

Results

The number of glomeruli per biopsy ranged from 1 to 72. Human annotators showed good agreement in the number and area of glomeruli (R2 = 0.75 and 0.66, respectively). Classification reached an overall machine-human accuracy in cross-validation of 94% on a dataset with an equal number of positive and negative examples. However, classification on full biopsies (with more negative than positive areas) had a low precision (45%) with an overall F1 score (harmonic mean of precision and recall) of 51%. In contrast to classification, segmentation applied to entire biopsies identified glomeruli with zero false positives and an overall F1 score of 99%. Neither model’s performance was altered by changes in stain intensity or the scanner used for imaging.

Conclusion

Glomeruli can be detected with high accuracy in renal biopsies using deep learning techniques; image segmentation worked better than image classification. These computational models can be leveraged to improve and automate enumeration of glomeruli, and can be adapted for classification of glomerular morphology.

Funding

  • NIDDK Support