Abstract: SA-PO711
PatchSorter Enables Efficient Digital Glomerular Classification
Session Information
- Glomerular Diseases: Clinical, Outcomes, Trials - III
November 05, 2022 | Location: Exhibit Hall, Orange County Convention Center‚ West Building
Abstract Time: 10:00 AM - 12:00 PM
Category: Glomerular Diseases
- 1303 Glomerular Diseases: Clinical‚ Outcomes‚ and Trials
Authors
- Ambekar, Akhil, Duke University Department of Pathology, Durham, North Carolina, United States
- Talawalla, Tasneem, Case Western Reserve University Department of Biomedical Engineering, Cleveland, Ohio, United States
- Wang, Bangchen, Duke University Department of Pathology, Durham, North Carolina, United States
- Cassol, Clarissa Araujo, Arkana Laboratories, Little Rock, Arkansas, United States
- Madabhushi, Anant, Case Western Reserve University Department of Biomedical Engineering, Cleveland, Ohio, United States
- Barisoni, Laura, Duke University Department of Pathology, Durham, North Carolina, United States
- Janowczyk, Andrew, Case Western Reserve University Department of Biomedical Engineering, Cleveland, Ohio, United States
Background
Quantification of segmental (SS) and global glomerulosclerosis (GS) has diagnostic and prognostic relevance in kidney disease, which entails laboriously assigning SS, GS, and non-GS/SS labels to individual glomeruli. While deep learning (DL) based tools can automate this task, they require large amounts of categorized glomeruli for training. PatchSorter (PS), an open-source tool (patchsorter.com), can facilitate review and label assignment of glomeruli by grouping those with similar presentational characteristics in a 2-dimensional plot. Within this plot, as PS receives user feedback, separation between categories iteratively increases to facilitate bulk labeling, improving labeling efficiency. This study compares glomerular labeling efficiency of PS versus an un-DL-aided approach, QuickReviewer (QR).
Methods
3446 segmented glomeruli from 241 NEPTUNE PAS whole slide images previously manually categorized as SS, GS, and non-SS/GS were uploaded in PS and QR for labeling by a pathologist. Labels per second (LPS) were calculated for both approaches. Concordance between PS labels and ground truth was calculated.
Results
All glomeruli were labeled in 4260 seconds using PS (0.808 LPS), vs 712 glomeruli in 1800 seconds using QR (0.395 LPS), yielding a 105% speed improvement for PS over QR. Concordance of PS labels with manual categorization was 95%, suggesting efficiency improvements did not come at the cost of labeling fidelity.
Conclusion
PS is a robust and efficient tool for glomerular classification with potential to aid in overcoming manpower limitations in generating large cohorts of labeled digital kidney biopsies.
Funding
- NIDDK Support