"Endomyocardial biopsy (EMB) screening represents the standard of care for detecting allograft rejections after heart transplant," Dr. Faisal Mahmood of Brigham and Women's Hospital in Boston and colleagues note in Nature Medicine. "Manual interpretation of EMBs is affected by substantial interobserver and intraobserver variability, which often leads to inappropriate treatment with immunosuppressive drugs, unnecessary follow-up biopsies and poor transplant outcomes."
The new AI system, called Cardiac Rejection Assessment Neural Estimator (CRANE), is designed to be used in tandem with clinical assessments to establish an accurate diagnosis quickly and reduce disagreements regarding the presence and severity of rejection.
"The model we present is just the first step towards showing that AI-driven assessment of pathology images from heart transplant biopsies is possible," Dr. Mahmood told Reuters Health by email. "Going forward, we are integrating the patient's entire electronic medical record besides the pathology image to potentially improve the model. We are also enhancing the generalizability of the model by including more patient cases for model training, and collaborating with additional institutions to further validate the models."
CRANE is a deep learning-based AI system for automated assessment of gigapixel whole-slide images obtained from EMBs. The team trained CRANE for detection, subtyping, and grading of transplant rejection using thousands of pathology images from over 1,300 heart biopsies from Brigham.
To assess the model's performance, the team curated a large dataset from the US, as well as from test cohorts from Switzerland and Turkey, thereby including large-scale variability across populations, sample preparations and slide scanning instrumentation.
The model detected allograft rejection with an area under the receiver operating characteristic curve (AUC) of 0.962; assessed the cellular and antibody-mediated rejection type with AUCs of 0.958 and 0.874, respectively; detected Quilty B lesions - benign mimics of rejection - with an AUC of 0.939; and differentiated between low- and high-grade rejections with an AUC of 0.833.
A human reader study showed that CRANE's performance was non-inferior to conventional assessment and reduced both interobserver variability and assessment time.
Dr. Mahmood said, "These models can potentially...have an impact on improving patient outcomes. The biggest challenge is that most pathology departments in hospitals around the country still use glass slides under a microscope to made diagnostic assessments. There needs to be a broad transition from using physical glass slides to digitally scanned images and then we will see AI for pathology much more broadly applicable."
Dr. Alex Reyentovich, medical director of the heart transplant program at NYU Langone Health in New York City, commented on the study in an email to Reuters Health, "In light of the subjectivity of pathologic assessment of cardiac rejection, this is a welcome potential tool that could bring standardization for the purpose of clinical care and answering research questions."
"Moving away from a single reader assessment of rejection (with) a more standardized method with less inter-observer variability is a step in the right direction," he said. "We need well done clinical trials to make sure that this strategy translates into improvement in patient outcomes, and not just increased cost of care."
SOURCE: https://go.nature.com/3DrcVn6 Nature Medicine, online March 21, 2022.
By Marilynn Larkin
Posted on
site created by:
© 2024 Medicom Medical Publishers. All rights reserved. Terms and Conditions | Privacy Policy