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: TH-OR44

A Deep-Learning Approach to Kidney Donor Biopsy Frozen Sections

Session Information

Category: Pathology and Lab Medicine

  • 1602 Pathology and Lab Medicine: Clinical

Authors

  • Davis, Richard, Duke University School of Medicine, Durham, North Carolina, United States
  • Li, Xiang, Duke University, Durham, North Carolina, United States
  • Souma, Nao, Duke University School of Medicine, Durham, North Carolina, United States
  • Wang, Zehan, Duke University, Durham, North Carolina, United States
  • Xu, Yuemei, Nanjing University Medical School Affiliated Nanjing Drum Tower Hospital, Nanjing, Jiangsu, China
  • Sotolongo, G., Duke University School of Medicine, Durham, North Carolina, United States
  • Bell, Jonathan, Duke University School of Medicine, Durham, North Carolina, United States
  • Ellis, Matthew Jay, Duke University School of Medicine, Durham, North Carolina, United States
  • Howell, David Noble, Duke University School of Medicine, Durham, North Carolina, United States
  • Shen, Xiling, Duke University, Durham, North Carolina, United States
  • Lafata, Kyle, Duke University, Durham, North Carolina, United States
  • Barisoni, Laura, Duke University School of Medicine, Durham, North Carolina, United States
Background

Pre-implant assessment of donor kidney biopsies to determine allograft viability is often performed by non-renal pathologists, and carries limited accuracy and reproducibility. The purpose of this work is to develop a deep learning (DL) method for the classification of relevant histologic primitives from donor biopsies as an aid tool to pathologists. Non-sclerotic and sclerotic glomeruli were selected to test this approach.

Methods

A total of 268 frozen sections stained with hematoxylin and eosin (H&E) from cadaveric donor kidney biopsies (128 performed at Duke and 140 at outside institutions) were scanned into whole slide images (WSI) at 40x (Leica Biosystems AT2). Duke WSIs were divided at the patient level into training and validation cohorts (0.8:0.2) and non-Duke WSIs were used as testing dataset. QuPath was employed to manually annotate non-sclerotic (22767) and sclerotic glomeruli (1366). A 9-layer convolutional neural network (CNN), based on the common U-NET architecture, was developed in Python, using randomly selected 256x256 patches from WSI, and image augmentation to boost generalization performance. CNN hyper-parameters were tuned via cross-validation. The CNN’s performance was quantified based on the Dice Similarity Coefficient (DSC) between the predicted and ground-truth annotations.

Results

For non-sclerotic glomeruli, the average DSC for train, validation and testing was 0.93, 0.91, 0.90 respectively. The F1, Recall, and Precision for testing was 0.93, 0.96, 0.90 respectively. For sclerotic glomeruli, the average DSC for train, validation and testing was 0.89, 0.87, 0.83 respectively. The F1, Recall, and Precision for testing was 0.87, 0.93, 0.81 respectively. The CNN had higher performance in the regions of high glomerular density and occasionally, outperformed the pathologists in glomerular detection. Lower model performance was observed in the presence of image artifacts and in regions of low glomerular density.

Conclusion

DL applied to image analysis may help standardize and improve accuracy and reproducibility of quantification of histologic primitives in kidney frozen sections, enabling the establishment of synergistic machine-human protocols that can be deployed in clinical practice. The development of DL-segmentation of other relevant histologic primitives is in process.

Funding

  • Private Foundation Support