A comparative deep learning approach using foundation models for identification of acute cellular rejection in endomyocardial biopsies of heart transplant recipients

Introduction Cardiac allograft rejection (CAR) is a major cause of graft failure, and accurate detection is crucial for timely intervention. Current histopathological evaluation of endomyocardial biopsy (EMB) samples relies on manual inspection, which is time-consuming and subject to interobserver variability. In recent years, deep learning has been increasingly explored to assist in automating the identification of lymphocytic infiltrates in digital pathology. As several foundational models and neural architectures are now readily available, it is necessary to benchmark their performance in specific clinical tasks. This study evaluates and compares existing state-of-the-art deep learning methods to identify the most effective and reliable approach for assessing cardiac rejection. Purpose This study aims to develop a deep learning-based approach for detecting and grading lymphocyte infiltration in EMB to improve the diagnosis of cardiac allograft rejection. By leveraging digital pathology, we seek to enhance diagnostic accuracy and reduce interobserver variability. Additionally, we aim to demonstrate the effectiveness of AI in this task by evaluating and comparing different state-of-the-art foundation models and deep learning architectures, highlighting the potential for AI-driven solutions even in resource-constrained settings. Methods We analyzed 976 histological slides from 251 patients, all scanned manually to ensure high-quality digitization. To extract meaningful features from the samples, we employed and compared two foundational models specifically designed for digital pathology: UNI and its successor, UNIv2. To process these features and aggregate information across multiple slides per patient, we tested and compared multiple approaches, including ABMIL and Convolutional Neural Networks. This comparative analysis allowed us to identify the mostrobust strategy for rejection diagnosis. To assess performance, we used AUC-ROC and accuracy as our primary metrics. Results On the test set, our best-performing model configuration achieved an average AUC-ROC of 0.89 and an average accuracy of 0.80. However, the system shows some limitations, as samples including even partial features of prior “biopsy sites” can exhibit patterns resembling lymphocytic infiltrates associated with rejection, occasionally leading to false positives. Conclusion Our results demonstrate the potential of deep learning in improving the diagnosis of cardiac allograft rejection by enabling an automated and consistent assessment of lymphocyte infiltration in EMB samples. By comprehensively evaluating foundation models (UNI and UNIv2) paired with various ABMIL and CNN architectures, we achieved strong performance despite the limited amount of available data. These results suggest that while the models effectively capture lymphocytic infiltration patterns, further refinements are needed to mitigate artifacts from “biopsy sites” and improve overall specificity and diagnostic reliability.