AI in the detection of acute heart transplant rejection

Background/Introduction

Endomyocardial biopsy (EMB) is the current gold standard for detecting acute heart transplant rejection [1]. However, it is invasive, samples only a small myocardial region, shows low interobserver agreement among pathologists in detecting rejection and may miss early-stage acute rejection [2]. Cardiovascular magnetic resonance (CMR) with quantitative T1 and T2 mapping offers non-invasive tissue characterization, and artificial intelligence (AI) may further improve diagnostic performance. Limited dataset sizes in transplant populations require advanced preprocessing and augmentation strategies to enable robust deep-learning (DL) model training.

Purpose

To develop and evaluate a DL based classification pipeline for detecting heart transplant acute rejection using quantitative T1 and T2 parametric CMR maps.

Methods

Fifty-seven pediatric and adult heart transplant patients (245 CMR studies; 1–7 scans per patient) were included in the study [3]. The study population was split into training and testing populations with 90/10% share, respectively. The training population consisted of 50 patients and 216 CMR studies, while the testing population consisted of 7 patients and 29 CMR studies.

The T1 mapping labeling approach used a hybrid-criteria model: For pediatric studies, cases were labeled as positive if the average T1 value exceeded 1092 ms across 11 heart segments. For adult studies, cases were labeled as positive if the average T1 value was above 1065 ms across 12 heart segments.

Native T1 and T2 mapping were performed in three short-axis slices using 1.5T scanners. Parametric T1 and T2 maps were computed using vendor-independent algorithms, followed by segmentation of the left ventricle with a dedicated CNN model. Data coverage was augmented with preprocessing, including scaling, rotation, translation and region of interest masking. Special emphasis was placed on decreasing the effect of patient and vendor specific technical variation.

The applied DL model had six inputs, three for each T1 and T2 map in basal, mid, and apical slice locations. Each of the inputs was input to a DenseNET network. The DenseNET layers were interconnected contrast-wise to maximize training power of the convolutional part of the model. The predicted classification label was generated by the final fully connected layer. Finally, performance of the model was assessed using area under the ROC curve (AUC).

Results

The DL model predicting acute transplant rejection achieved excellent diagnostic performance, with AUC of 0.977.

Conclusions

A DL model trained on quantitative T1 and T2 CMR maps can reliably detect acute heart transplant rejection. The combination of advanced preprocessing, segmentation, and extensive augmentation enables model training despite limited patient numbers. CMR-based AI support tools have the potential to reduce reliance on invasive EMB and enhance non-invasive rejection monitoring.Fig 1.Applied network model.Fig 2.ROC.