Kidney Week

Abstract: SA-PO003

Computational Characterization of Lymphocytic Inflammation on Digital Kidney Biopsies

Session Information

• Bioengineering
  November 05, 2022 | Location: Exhibit Hall, Orange County Convention Center‚ West Building
  Abstract Time: 10:00 AM - 12:00 PM

Category: Bioengineering

• 300 Bioengineering

Authors

• Li, Xiang, Duke University, Durham, North Carolina, United States

Xiang Li,

• Sotolongo, Gina, Duke University Hospital, Durham, North Carolina, United States

Gina Sotolongo,

• Hodgin, Jeffrey B., University of Michigan, Ann Arbor, Michigan, United States

Jeffrey B. Hodgin,

• Mowery, Yvonne M., Duke University, Durham, North Carolina, United States

Yvonne M. Mowery,

• Holzman, Lawrence B., University of Pennsylvania, Philadelphia, Pennsylvania, United States

Lawrence B. Holzman,

• Mariani, Laura H., University of Michigan, Ann Arbor, Michigan, United States

Laura H. Mariani,

• Chen, Yijiang, Case Western Reserve University, Cleveland, Ohio, United States

Yijiang Chen,

• Madabhushi, Anant, Case Western Reserve University, Cleveland, Ohio, United States

Anant Madabhushi,

• Zee, Jarcy, University of Pennsylvania, Philadelphia, Pennsylvania, United States

Jarcy Zee,

• Janowczyk, Andrew, Case Western Reserve University, Cleveland, Ohio, United States

Andrew Janowczyk,

• Barisoni, Laura, Duke University, Durham, North Carolina, United States

Laura Barisoni,

• Lafata, Kyle, Duke University, Durham, North Carolina, United States

Kyle Lafata,

Background

Dispersion patterns of inflammatory cells in kidney tissue are hypothesized to be associated with patient outcomes, yet remain challenging to quantify visually in a standardized fashion. This study developed a deep learning (DL) model to automatically identify lymphocytic inflammation on H&E images.

Methods

18 FFPE kidney specimens with moderate to severe inflammation were H&E stained and scanned at 40x, then rescanned after IHC staining for CD3/CD20 (T+B lymphocytes). After co-registering H&E and IHC, lymphocytes were identified as IHC-positive cells, with all other cells being negative. A Hover-Net DL model was trained to segment these IHC-positive cells using 6 whole slide images (WSIs) and tested on 12 WSIs. Model generalization was evaluated on NEPTUNE WSIs (visual inspection). Cell graphs were calculated to characterize the topology of the lymphocytic microenvironment. The predicted topology on H&E was compared to the measured topology on IHC via the structural similarity index measure (SSIM).

Results

The training and testing datasets included 22,732 nuclei (12,618 lymphocytes) and 7,984 nuclei (3,712 lymphocytes), respectively. The model precision, recall, f1, and AUC were 0.69, 0.77, 0.73, and 0.74, respectively. The topology-pattern level similarity was high, with mean SSIM of 0.82±0.05 between H&E and IHC WSIs. In the NEPTUNE dataset, the DL model correctly differentiated lymphocytic from non-lymphocytic inflammation.

Conclusion

Co-registering restained H&E and IHC WSIs provides an efficient way of generating large quantities of training material for DL. Here, our model accurately recapitulated lymphocyte topology patterns using only H&E. These patterns will subsequently be tested for outcome prediction in glomerular diseases.

Funding

• NIDDK Support