Prediction of heart transplant rejection from routine pathology slides with self-supervised deep learning

Abstract

Background and Aims One of the most important complications of heart transplantation is organ rejection, which is diagnosed on endomyocardial biopsies by pathologists. Computer-based systems could assist in the diagnostic process and potentially improve reproducibility. Here, we evaluated the feasibility of using deep learning in predicting the degree of cellular rejection from pathology slides as defined by the International Society for Heart and Lung Transplantation (ISHLT) grading system. Methods We collected 1079 histopathology slides from 325 patients from three transplant centers in Germany. We trained an attention-based deep neural network to predict rejection in the primary cohort and evaluated its performance using cross validation and by deploying it to three cohorts. Results For binary prediction (rejection yes/no) the mean Area Under the Receiver Operating Curve (AUROC) was 0.849 in the cross-validated experiment and 0.734, 0.729 and 0.716 in external validation cohorts. For a prediction of the ISHLT grade (0R, 1R, 2/3R), AUROCs were 0.835, 0.633 and 0.905 in the cross-validated experiment and 0.764, 0.597, 0.913, and 0.631, 0.633, 0.682, and 0.722, 0.601, 0.805 in the validation cohorts, respectively. The predictions of the AI ​​model were interpretable by human experts and highlighted plausible morphological patterns. Conclusions We conclude that artificial intelligence can detect patterns of cellular transplant rejection in routine pathology, even when trained on small cohorts.

Regression-based Deep-Learning predicts molecular biomarkers from pathology slides

“…Traditional DL methods have limited the analysis of many biomarkers, including HRD and gene expression signatures, which are continuous values, by categorizing them into discrete classes. Our study provides direct evidence that regression networks, such as the CAMIL regression method described in this study, which builds on recent work using attention-based multiple instance learning and self-supervised pre-training of the feature extractor 18 , 20 , 33 , outperforms traditional classification and regression networks in predicting these biomarkers. This approach unlocks a key clinical application area for pathology-based biomarker prediction.…”

“…These regions may, however, contain less relevant tissues such as connective tissue or fat, which might not contribute to biomarker predictability 14 . To address this issue, attention-based multiple instance learning (attMIL) is the predominant technical approach that is currently used 15 – 18 . To implement this strategy, feature vectors are first extracted from pre-processed tiles.…”

Cardiac Allograft Vasculopathy: Challenges and Advances in Invasive and Non-Invasive Diagnostic Modalities

“…The role of artificial intelligence (AI) in the early detection of CAV is not yet established; however, it is an area of ongoing research and innovation [ 120 ]. In a review article about the current state of artificial intelligence in cardiac transplantation, Goswami suggests that AI could be used to analyze image segmentation of cardiac biopsies, as well as genomic and proteomic data, to identify new factors that influence the development and progression of CAV [ 121 ].…”

A Preliminary Study Assessing a Transfer Learning Approach to Intestinal Image Analysis to Help Determine Treatment Response in Canine Protein-Losing Enteropathy

“…To utilize whole slide images (WSIs) for training the machine learning model, we adopted the Aachen Protocol (https://zenodo.org/record/3694994; accessed on 11 April 2023) to extract tiles from WSIs and prepared images for machine learning. Aachen protocols were applied in the development of various deep learning models, such as those used for predicting heart transplant rejection [14] and detecting Epstein-Barr virus and microsatellite instability [15], as well as predicting genetic alterations based on HE slides of gastric cancer patients [16]. We employed QuPath v0.4.3 [17] using tessellate WSIs into tiles of 512 × 512 pixels at 0.22 µm per pixel.…”