Heart transplant

AI system helps spot signs of heart transplant rejection

An artificial intelligence (AI) tool can help identify heart transplant rejection and estimate its severity, results from a pilot study suggest.

The Cardiac Rejection Assessment Neural Estimator (CRANE) simultaneously addresses the detection, subtyping and grading of allograft rejection in H&E stained whole slide images of endomyocardial biopsy specimens and is intended to be used in conjunction with the heart transplant team to more quickly and accurately diagnose rejection.

“Our retrospective pilot study demonstrated that the combination of artificial intelligence and human intelligence can improve expert agreement and reduce the time needed to assess biopsies,” said lead author Faisal Mahmood, PhD, Department of Pathology, Brigham and Women’s Hospital, Boston, in a press release. .

The study was published in the March issue of natural medicine.

Improved standard of care

Endomyocardial biopsy screening is the standard of care for detecting cardiac allograft rejection, but manual interpretation of surveillance endomyocardial biopsies remains a challenge, the authors note. Experts often disagree on whether or not the patient rejects the allograft and how severe the rejection is when present.

Overestimating rejection can lead to patient anxiety, overtreatment, and unnecessary follow-up biopsies, while underestimating can lead to treatment delays and poorer outcomes.

“CRANE is an AI model that can act as an assistive tool to reduce this observer variability,” Mahmood said. lecoeur.org | Medscape Cardiology.

“Although the final assessment remains subjective, the CRANE model provides experts with an AI-based prediction, as well as its confidence in achieving those predictions,” he said.

“To inspire further confidence, the model also highlights the regions of the image from which it makes the prediction,” he added.

CRANE was trained on more than 5,000 gigapixel whole-slide images from nearly 1,700 patient endomyocardial biopsy specimens taken from Brigham and Women’s Hospital. Model performance was determined using a separate large dataset from the United States, as well as independent test cohorts from Turkey and Switzerland.

“These independent international test sets were deliberately constructed to reflect the wide variability of data present in populations and medical centers, as they used different biopsy protocols, slide preparation and staining mechanisms, and scanner vendors, which are all known contributors to image variability,” the researchers note in their report.

The results show that CRANE detects allograft rejection, with an area under the receiver operating characteristic curve (AUC) of 0.962; assesses type of cellular and antibody-mediated rejection with AUCs of 0.958 and 0.874, respectively; detects Quilty B lesions (benign imitations of rejection) with an AUC of 0.939; and differentiates low- and high-grade rejection with an AUC of 0.833.

pass glass Slides to digital scans

CRANE shows “non-inferior performance to conventional assessment and reduced inter-observer variability and assessment time,” the researchers report.

Mahmood said further studies and clinical trials are needed to establish the overall effectiveness of evaluating AI-assisted endomyocardial biopsy for cardiac allograft rejection and its potential to improve transplant outcomes. cardiac.

He noted that the biggest obstacle to the broad applicability of these AI models is that most hospital pathology departments nationwide still use glass slides under a microscope to perform diagnostic assessments.

“First, there needs to be a broad transition from using physical glass slides to digitized images, and then we’ll see AI for pathology much more widely applicable,” he said.

Reached for comment, Preethi Pirlamarla, MD, a heart failure and transplant cardiologist at Mount Sinai Queens in New York, said CRANE is a “very promising” tool that uses deep learning to enable better evaluation of endomyocardial biopsies.

“One of the drawbacks of endomyocardial biopsies is that there is very high inter-reader and inter-observer variability, as well as sampling issues that can confound the results. This tool could serve as another layer to assess the cardiac allograft rejection”, Pirlamarla told lecoeur.org | Medscape Cardiology.

One of the main strengths of the study is the use of three different cohorts from the United States, Turkey and Switzerland, with different protocols. “It really speaks to the stability of the program, even when you use it in different countries and different protocols,” Pirlamarla said.

This study “should serve as a springboard for other larger-scale trials,” she added.

This work was supported in part by the Brigham and Women’s (BWH) Presidential Fund, internal funds from BWH and Massachusetts General Hospital, the National Institutes of Health, and the National Science Foundation. Mahmood and Pirlamarla report that no conflicts of interest.

NatureMed. 2022;28:575-582. Full Text