Predicting heart rejection using histopathological whole-slide imaging and deep neural network with dropout

Abstract

Cardiac allograft rejection is one major limitation for long-term survival for patients with heart transplants. The endomyocardial biopsy is one gold standard to screen heart rejection for patients that have heart transplantation. However, manual identification of heart rejection is expensive and time-consuming. With the development of imaging processing techniques and machine learning tools, automatic prediction of heart rejection using whole-slide images is one promising approach to improve the care of patients with heart transplants. In this paper, we first develop a histopathological whole-slide image processing pipeline to extract features automatically. Then, we construct deep neural networks with and without regularization and dropout to classify the patients into nonrejection and rejection respectively. Our results show that neural networks with regularization and dropout can significantly reduce overfitting and achieve more stable accuracies.

Automated Classification of Acute Rejection from Endomyocardial Biopsies

“…Additionally, the heterogeneity of morphological phenotypes in ACR and AMR, combined with the limited availability of localized feature annotations, presents a significant challenge to patch-labeling based approaches. Prior studies have proposed weakly supervised machine learning algorithms to predict rejection status from whole slide images, combined with attention modeling or whole-slide aggregation of patch-level features [10,27,33].…”

“…There is limited work on the automatic prediction of heart transplant rejection using endomyocardial biopsies due to the difficulty in obtaining labeled WSI datasets for research. Tong et al implemented a deep neural network with regularization and dropout to classify patients into rejection and non-rejection [27]. The wholeslide images were separated into regions of interest (ROIs) which were further broken into patches.…”

“…Next, each WSI was tiled into 256 × 256 × 3 patches with a 50% overlap between neighboring patches in both the vertical and horizontal axes. This patch area was chosen after examining many options at different WSI magnification levels and concluding that patches of size 256 × 256 × 3 at 10X magnification was consistent with detecting rejection according to previous studies and reviews on the subject [10,19,27,33]. Patch coordinates were then transformed into thumbnail coordinates and used to calculate the percent tissue coverage for each patch from the tissue mask thumbnail.…”

Improved Prediction on Heart Transplant Rejection Using Convolutional Autoencoder and Multiple Instance Learning on Whole-Slide Imaging

“…Barker et al [4] extracted a total of 227 color and shape-based features from brain tumor pathological WSI tiles prior to WSI-level classification using the elastic net and weighted voting. Tong et al [5] extracted 461 features such as nuclear density and grey level co-occurrence matrix (GLCM) from heart transplant WSIs before merging and classifying these features using deep neural networks with dropout. Dooley et al [6] pointed out that additional object-level and pixel-level features could be extracted from the heart transplant WSIs.…”

Prediction of heart transplant rejection using histopathological whole-slide imaging

“…Manually screening heart transplant rejection can be costly and time-consuming [3]. Distinguishing between ACR and AMR using endomyocardial biopsies, and the grades of each type of rejection, can be subjective due to a lack of agreement among pathologists of histologic diagnostic criteria.…”