Multimodal AI based Assessment of Renal Allograft Biopsies

Principal Investigator: Faisal Mahmood, Astrid Weins

Authors: Sharifa Sahai
Lay Abstract
Scientific Abstract

There are over 100,000 kidney transplants annually with over 24,000 transplants in the United States alone. Manual assessment of renal biopsies suffers from large inter- and intra-observer variability. Such variability can have dire consequences ranging from under and over treatment to partial or full transplant rejection or even death. Moreover, renal allograft assessment is a complicated process involving multiple tissue stains and several modalities and requires the expertise of a renal pathologist. Thus, we propose a multimodal AI based assessment of renal allograft biopsies for renal allograft rejection. We utilize multimodal fusion to account for information from Hematoxylin and eosin, Periodic acid Schiff, Masson trichrome and Jones silver wholes slide images for overall transplant assessment. The multiple instance learning module assigns attention scores to each tissue patch through ranking the importance of each patch to the label provided for the slide. The attention scores can be translated to WSI attention heatmaps reflecting relevance of each biopsy region towards the model predictions, which can be used for validation and interpretation by pathologists. Our preliminary results are promising with regards to the tasks of active cell mediated rejection, antibody mediated rejection and interstitial fibrosis and tubular atrophy classification and interpretability.

Clinical Implications
Once deployed this model could be a great asset to both clinical and technical methods driven research. The heat maps corresponding to the rejection classes can be used to help diagnose renal allograft rejection.

If the PDF viewer does not load initially, please try refreshing the page.

Comments are closed.