BANFF-AID: Banff Automated Nephrology Feature Framework - Artificial Intelligence Diagnosis

Project Summary (30 lines) Banff Lesion Scores are utilized globally for assessing the condition of allografts, particularly donor kidneys. They are also used to predict the overall health of the transplanted organ. However, visual human assessment, coupled with imprecise definitions, leads to inter-observer and intra-observer variability. This subjectivity in

evaluation often results in conservative decisions, causing potentially viable donor organs to be discarded, representing missed opportunities. Furthermore, detailed visual assessment by pathologists is time consuming, and its monotony can lead to increased errors and worker burnout. Increasing automation of this analysis will

alleviate both of these challenges. Our project, BANFF-AID, Banff Automated Nephrology Feature Framework - Artificial Intelligence Diagnosis, aims to transform and significantly improve this process by introducing artificial intelligence (AI)-driven automated score generation and a user-centric web-based workflow for kidney biopsies. Through

the integration of custom-built AI models that extract essential features from biopsy slides, our system delivers accurate and continuous (non-discretized) Banff Lesion Scores. Our approach of using AI-produced image segmentation of features of interest ensures that the results are explainable and the method is transparent.

This approach also empowers pathologists to intervene and make corrections to the segmentation in case of incorrect scoring due to inaccurate feature segmentation. Our software framework builds upon previous efforts supported by various NIH grants, providing a strong foundation. In Phase I, our proof of concept development will focus on renal tissues featuring arteriosclerosis,

glomerulosclerosis, and IFTA (interstitial fibrosis and tubular atrophy). We will be working with H&E (hematoxylin and eosin) stained frozen sections using more precise definitions than those employed in routine histopathology practice. Moving into Phase II, our goal is to expand the scope to compute multiple Banff Lesion

Scores from multiple slides and multiple stains. This comprehensive approach will reduce the burden on the pathologists, allowing them to focus on interpreting measurements and providing diagnosis. 1