Abstract: FR-PO1012
Deep-Learning Based Pathological Assessment in Frozen Procurement Kidney Wedge Biopsies: An Independent Validation Study
Session Information
- Transplantation: Basic
October 25, 2024 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Transplantation
- 2101 Transplantation: Basic
Authors
- Yi, Zhengzi, Icahn School of Medicine at Mount Sinai, New York, New York, United States
- Zhang, Weijia, Icahn School of Medicine at Mount Sinai, New York, New York, United States
Background
In order to minimize observer variability and avoid inappropriate discard pre-transplant, a deep-learning based pathological assessment pipeline in frozen procurement kidney needle biopsies has been developed and was shown to be able to accurately capture normal/sclerotic glomeruli, arteries/arterial intimal fibrosis regions and tubules. A composite Kidney Donor Quality Score (KDQS) was derived and used in combination with clinical factors to predict graft loss or assist organ utilization. However the performances of pipeline and graft loss model in frequently performed wedge biopsies were not thoroughly investigated.
Methods
We processed an independent cohort of procurement kidney wedge biopsies (n=786) which were mostly transplanted in 2019-2021 and were provided by Gift of Hope OPO center. Clinical data were obtained from OPTN under institutional IRB approval. Glomerulosclerosis grade was the only available pathologists’ score and was used to correlate with digital feature Sclerotic Glomeruli%. Performance of pre-trained 1-year graft loss model using KDQS and clinical factors: cold ischemic time (CIT), use of pump and induction therapy was also evaluated.
Results
The digital feature Sclerotic Glomeruli% was strongly correlated with pathologists’ glomerulosclerosis grade (R=0.63, p=1.8e-62). Sclerotic Glomeruli% (6m: p=9.6e-07, 12m: p=4.4e-07), Arterial Intimal Fibrosis% (6m: p=0.003, 12m: p=4.8e-05), and Interstitial Space Abnormality% (6m: p=0.003, 12m: p=0.016) were all significantly correlated with post-transplant eGFR. The KDQS was significantly associated with graft loss (p=5.4e-04, HR=1.39). Both individual and composite digital features were superior to the pathologists’ glomerulosclerosis grade in association with graft outcomes. By applying the pre-trained 1-year graft loss model to wedge biopsy cohort using KDQS and clinical factors, we obtained an AUC of 0.67, which surpassed the model performance using clinical factors (AUC=0.53) or KDPI (AUC=0.57) alone. KDQS>=7 identified 5 kidneys could have been discarded within which 1 patient lost graft at day 122, 3 patient had 12m eGFR<=32, the rest patient had moderate eGFR (47) at 6m.
Conclusion
This study further validated the previously trained needle biopsy based method in wedge biopsies, therefore expended the utility of our deep-learning pipeline in clinical practice.
Funding
- Private Foundation Support