In this project, we aim to develop a computer-aided diagnosis (CAD) tool to help doctors with objective and quantitative evaluation of heart rejection. Heart rejection is one of the leading causes of death for children with a heart transplant. To diagnose heart rejection, doctors need to examine the pathological images of heart biopsy samples manually. However currently, this process is time-consuming and subjective.

With the development of medical imaging informatics and artificial intelligence, we want to assist the doctors by providing automated feature extraction and objective decision support. Thus, (1) we have developed an advanced AI analytic such as deep-learning model (e.g. convolutional atoencoders) to extract pathological image features for describing different cardiac tissue rejection categories; (2) we have developed a method to aggregate these features; and (3) we have obtained local annotation of features needed for deep residual convolutional network in whole slide image quantification.

Because the current diagnosis of heart rejection does not necessarily reflect the clinical outcomes of the patient, we want to improve the correlation between the diagnosis and the clinical outcomes by designing a novel scoring system that can indicates the severity of heart rejection better. Thus, we will continually develop this tool, i.e., computer aided diagnosis systems (CADS) for cardiologist and pathologist to improve diagnosis accuracy. We plan (A) to develop an adaptive sampling approach for accurate localization of heart rejection regions on whole-slide images; (B) to develop a weakly supervised learning model to learn from wholeslide images without local annotations; and (C) to develop a novel scoring system for heart rejection diagnosis. Ultimately, we hope this new tool with novel scoring system can improve the clinical care for the children with heart rejection and eventually save their lives.

Heart transplant rejection must be quickly and accurately identified to optimize anti-rejection therapies and prevent organ loss. Expert evaluation of endomyocardial biopsies is labor-intensive, and prone to human bias, and suffers from low inter-rater agreement. Additionally, the increased utility of digital pathology for biopsy examination has exacerbated the need for additional image quality control. To meet these challenges, we developed a novel transplant rejection detection pipeline which automatically identifies histology slides in need of rescanning and highlights biopsy regions showing potential signs of rejection. Our system leverages a fast and effective automated patch-level quality filter as well as state-of-the-art feature extraction techniques to provide quality whole-slide level labeling of early rejection signs. We successfully identified digital pathology images with poor image quality and leveraged this quality gain to improve our novel weakly-supervised learning model leading to significant transplant rejection classification performance of AUC: 70.12 (±20.74) %.

With heart transplant, hundreds of children with severe congenital or end-stage heart diseases are saved each year. In the United States, 507 of the 3551 heart transplants (14%) are performed for children (17 years old or under) in 2019. However, heart rejection is one of the major causes of death post heart transplant, and early and accurate identification of heart rejection is critical for timely and guided therapy to save the lives of these children. The gold standard for diagnosing heart rejection is pathological analysis of the heart tissues obtained by the endomyocardial biopsy (EMB). Because heart rejection cells are often patchy in each biopsy slide and pathologist’s scanning is heavily dependent on a particular “field of view”, the diagnosis for a given biopsy specimen can be subjective. Therefore, computational pathology is proposed to quantitatively analyze heart tissues to objectively diagnose heart rejection stages. The early diagnosis of heart rejection with tissue biopsy can be seen as a problem of “finding a needle in a haystack.” Most heart tissues obtained by the biopsy are healthy, while only a small portion of them are with rejection. The goal of this project is to utilize deep learning methodologies to automatically identify all these areas with rejection in the biopsy slide. The technology we are developing can pre-identify hot spots in a particular slide, thereby assisting pathologiest reading it. Additionally, we are improving the algorhithms to better grade the variations in the biopsy to fine tune our reporting. We have already published our recent work in academic journals and presented at top biomedical sciences conference