Antibody mediated rejection (AMR) occurs in 10-20% of cardiac transplant recipients and is undoubtedly associated with accelerated graft failure, and increased morbidity and mortality [1]. The incidence of AMR is expected to rise due to the increased number of pre-sensitized patients who have previously been supported by mechanical circulatory devices. AMR develops when recipient antibodies are directed against donor human leukocyte antigens (HLA) and possibly non-HLA on the endothelial layer of the allograft, inducing fixation and activation of the complement cascade [2].

To date, tissue diagnosis by endomyocardial biopsy remains the gold standard for pathologic AMR (pAMR), which is heavily relied upon by transplants cardiologists to correlate clinically significant AMR requiring plasmapheresis. The 2004 International Society for Heart and Lung Transplantation (ISHLT) revision of cardiac pAMR criteria details the two most important histologic criteria as 1) capillary injury with endothelial swelling and 2) intravascular macrophage accumulation. Ancillary studies to confirm these findings include immunofluorescence (IF) for deposition of immunoglobulin (IgG, IgM, and/or IgA) with complement (C1q, C4d, and/or C3d) on frozen section tissues, and CD68 staining of intracapillary macrophages or C4d staining of capillaries by immunohistochemistry (IHC) on paraffin-imbedded tissue. The 2013 ISHLT update expands these descriptions to include 1) accumulation of macrophages that distend the capillary and venule lumens and 2) enlarged endothelial nuclei and expanded cytoplasmic projections that either narrow or occlude the vascular lumens [3]. These descriptions were incorporated to clarify the histopathologic criteria for cardiac pAMR and reduce variability in the diagnosis.

The scoring system for pAMR includes both a histologic (pAMR-H) and immunopathologic component (pAMR-I). Diagnosing pAMR-H is based on evaluating cellular morphology by hematoxylin and eosin (H&E) staining and entails identifying the presence of the features described in the histologic ISHLT criteria. The recommended protocol for evaluation of the immunopathologic component on the endomyocardial biopsy include both frozen tissue for IF studies and paraffin IHC analysis to confirm deposition of complement on endothelium as well as infiltration of the intravascular macrophages. However, identifying pAMR-H, a qualitative morphologic assessment, can be much more challenging to diagnose than pAMR-I the immunologic component, and is associated with high inter-observer variability amongst even experienced cardiac pathologists [4]. Furthermore, unrecognized or misdiagnosis from inexperienced pathologists may lead to suboptimal treatment for transplant patients.

Detecting pAMR-H may be subtle and is most commonly mistaken for non-specific healing injury or the more frequent acute cellular rejection (ACR). Recent advances in digital pathology, such as the advent of whole slide scanners, faster computing power, increased and less expensive data storage options, and improved networking capabilities have allowed digital pathology to be applied to a wide variety of applications in the field of medicine. One such application that has gained wide traction in industry and academic settings is the application of machine learning algorithms in identifying disease processes using whole slide images (WSI) [5].

In particular, deep learning convolutional neural networks provide more accurate and robust network architectures than traditional machine learning approaches and have superior accuracy in image analysis applications [6], [7], [8]. Many image analyses tasks, such as segmentation, detection, counting, and tissue classification have been performed with greater accuracy than traditional machine learning methods and more accurately than humans. While much work has been in tissue classification in the area of neoplasia, there is a paucity of such investigations in the field of transplantation. We present here one of the first machine learning algorithms to our knowledge, developed to diagnose pAMR-H in cardiac transplant surveillance biopsies with a high degree of accuracy while distinguishing it from the more common form of rejection (ACR) and its benign histologic mimickers such as healing injury.