Automated Histology Lesion Interpretation in Kidney Transplant Biopsies Shows That Pathologists Often Deviate from Banff Guidelines
1University of Alberta, Edmonton AB, AB, Canada, 2Alberta Transplant Applied Genomics Centre, Edmonton, AB, Canada, 3Medical University of Vienna, Vienna, Austria, 4Medical University of Vienna, Edmonton AB, AB, Canada, 5Medical University of Warsaw, Warsaw, Poland, 6Pomeranian Medical University, Szczecin, Poland
Meeting: 2019 American Transplant Congress
Abstract number: A192
Keywords: Gene expression, Kidney, Kidney transplantation, Rejection
Session Information
Session Name: Poster Session A: Kidney Chronic Antibody Mediated Rejection
Session Type: Poster Session
Date: Saturday, June 1, 2019
Session Time: 5:30pm-7:30pm
Presentation Time: 5:30pm-7:30pm
Location: Hall C & D
*Purpose: Histologic diagnosis of kidney transplant biopsies requires an expert pathologist using Banff guidelines to interpret features. Some problematic guidelines require experts to use professional judgment. To explore variation (”noise”) in guideline applications, we created an “AutoBanff” algorithm that strictly applies Banff guidelines to recorded lesions.
*Methods: We studied 1679 prospective indication kidney transplant biopsies with lesion scores and clinical data (Clinical trials.gov #NCT01299168). Because Banff 2017 contains ambiguous elements, we used Banff 2015. An algorithm was developed using Banff guidelines, checked by expert pathologists, and the automated diagnoses compared to recorded histology (ExpertBanff) and molecular diagnoses (MMDx). Assigned diagnoses were based on a six-class model (ABMR, possible ABMR ‘pABMR’, TCMR, possible TCMR ‘pTCMR’, Mixed rejection, or No rejection ‘NR’). “Clear” discrepancies were between distinct classes (e.g. ABMR-NR), “boundary” discrepancies reflected ambiguity (e.g. pABMR-ABMR).
*Results: AutoBanff diagnoses compared to ExpertBanff (Table 1) disagreed in 439 biopsies (26%). Discrepancy was more frequent in molecularly abnormal biopsies; biopsies with scarring (ci2/3) or v-lesions>0; and biopsies with negative or ambiguous DSA or positive BK virus. Clear discrepancies represented 46% of discrepancies. In 53 clear discrepancies experts called NR, AutoBanff called 30 ABMR and 23 other. In 75 experts called ABMR, AutoBanff called 59 NR and 16 other. In 14 experts called TCMR, AutoBanff called 7 Mixed and 5 other. The commonest clear discrepancy was NR vs. ABMR (89/202); the commonest boundary discrepancy was NR vs. pTCMR (82/237). Of interest, ExpertBanff agreed more with MMDx than did AutoBanff (p=0.002), confirming that pathologists’ judgment added value.
*Conclusions: Histology lesions can be interpreted by a computerized algorithm. The 26% discrepancy between AutoBanff and ExpertBanff reflects the intrinsic noise in histology, which is concentrated in certain scenarios – e.g. DSA negative ABMR, scarred biopsies, v-lesions, and BK nephropathy – that suggest areas for special focus in refining Banff guidelines.
To cite this abstract in AMA style:
Madill-Thomsen KS, Reeve J, Bohmig G, Eskandary FA, Perkowska-Ptasinska A, Myslak M. Automated Histology Lesion Interpretation in Kidney Transplant Biopsies Shows That Pathologists Often Deviate from Banff Guidelines [abstract]. Am J Transplant. 2019; 19 (suppl 3). https://atcmeetingabstracts.com/abstract/automated-histology-lesion-interpretation-in-kidney-transplant-biopsies-shows-that-pathologists-often-deviate-from-banff-guidelines/. Accessed May 18, 2025.« Back to 2019 American Transplant Congress