Principal Investigator: Faisal Mahmood
Heart-rejection is a serious condition occurring in 30-40% of patients within one-year post-transplantation. Since most rejection appear without symptoms, cardiac biopsies are used for patient-monitoring and treatment-planning. The complexity of the biopsy microscopy image can be compared to Googlemaps view of the world, with different information captured at different magnification. This complexity significantly affects accuracy and time of manual biopsy inspections. Diagnostic inaccuracy often leads to over-or under-treatment with immunosuppressive drugs, unnecessary follow-up biopsies, and may lead to poor transplant outcomes. We present AI model, called CRANE, for automated assessment of cardiac biopsy images, which simultaneously address detection, subtyping, and grading of cardiac rejection. Concurrently, CRANE provides visual maps reflecting diagnostic relevance of each biopsy regions, enabling visual interpretation of the prediction and guidance for manual biopsy inspection. An independent reader study of five cardiac pathologists shows that the CRANE predictions are not inferior to human pathologists. Moreover, utilizing the assistance from the visual maps during manual biopsy assessments leads to lower inter-rater variability and the assessment time. This robust evaluation of the AI system paves the way for clinical trials to establish the efficacy of AI-assisted cardiac biopsy reads to improve heart transplant outcomes.
Surveillance endomyocardial biopsies (EMBs) are the standard-of-care for detecting rejection of a donor’s heart following cardiac transplantation. Histologic assessment of EMBs is, however, accompanied by substantial inter- and intraobserver variability, which often leads to over-or under-treatment with immunosuppressive drugs, unnecessary follow-up biopsies, and may lead to poor transplant outcomes. Here, we present a deep learning-based model, called CRANE, for automated assessment of whole-slide images (WSIs) of EMBs, which simultaneously address detection, subtyping, and grading of allograft rejection. Concurrently, CRANE provides WSI-level attention maps reflecting the diagnostic relevance of each biopsy region, enabling visual interpretation of the predictions and guidance for manual biopsy reads. To assess the model performance, we curated a large dataset from USA and independent test cohorts from Turkey and Switzerland representing large variability across populations, sample preparation, and slide scanners. An independent reader study of five cardiac pathologists shows that the CRANE predictions are not inferior to human pathologists. Moreover, utilizing the assistance from the attention maps during manual biopsy assessments leads to lower inter-rater variability and the assessment time. This robust evaluation of the AI system paves the way for clinical trials to establish the efficacy of AI-assisted biopsy assessment to improve heart transplant outcomes.
If the PDF viewer does not load initially, please try refreshing the page.
