https://orcid.org/0000-0003-1711-3098
This editorial refers to ‘An automated computational image analysis pipeline for histological grading of cardiac allograft rejection’†, by E.G. Peyster et al., on page .
Hypothetical workflow of cardiac allograft rejection diagnostics after implementation of recent developments. New non-invasive screening methods for rejection diagnostics are being developed, such as detection of donor-specific and non-HLA (human leucocyte antigen) antibodies, leucocyte gene expression profiling, and circulating donor-derived cell-free DNA (ddcfDNA). When a rejection is suspected, endomyocardial biopsy is performed. New tissue gene expression analysis methods can help improve the precision and accuracy of diagnosis. The microscope of the pathologist is replaced by a fully digital workflow where automated computational pathology such as CACHE can help pathologists to come to better and more standardized diagnoses.This figure was created using images from Servier Medical Art (https://smart.servier.com), which are licensed under a Creative Commons Attribution 3.0 Unported License.
For more than 40 years, the monitoring of acute cardiac allograft rejection has been done by serial invasive endomyocardial biopsies (EMBs) that are assessed by specialized cardiovascular pathologists through basic histology. This method started with the development of a bioptome by Philip Caves and Margaret Billingham in 1973 describing the first histological grading system.1 In 1990, a standardized international grading system for the histopathological diagnosis of acute rejection in EMB was adopted by the International Society for Heart and Lung Transplantation (ISHLT)2 and improved in later years. In this grading system, diagnostic criteria are provided for the two phenotypes of acute rejection: acute cellular rejection (ACR) and antibody-mediated rejection (AMR). Despite this effort to arrive at standardization, it is recognized that the reliability of EMB interpretation by pathologists is limited by high interobserver variability in discriminating true acute rejection from non-rejection effects, while the invasive nature of this method is another disadvantage.
In recent years, multiple non-invasive technologies have been developed that are candidates for monitoring acute cardiac allograft rejection3 (Graphical Abstract). Subclinical alterations in myocardial function can be measured by imaging techniques such as tissue Doppler imaging and speckle-tracking echocardiography. Activation of the immune system directed against the allograft can be screened by measuring complement-activating donor-specific and non-HLA (human leucocyte antigen) antibodies, leucocyte gene expression profiling, and circulating donor-derived cell-free DNA (ddcfDNA). During rejection episodes, allograft cells die and release short DNA fragments into the recipient’s circulation that can serve as biomarkers for acute cardiac allograft rejection. In a recent multicentre prospective cohort study, monitoring with ddcfDNA demonstrated excellent performance characteristics not only for ACR, but also for AMR which is more challenging to diagnose with basic histology.4 Interestingly, increases in ddcfDNA preceded both forms of acute rejection detected in biopsies. It is conceivable that in the near future these non-invasive techniques will be more widely used as screening methods, whereas EMBs are used to confirm the acute rejection.
In EMB diagnostics, gene expression analysis techniques are being developed that can support basic histology and improve the precision and accuracy of diagnoses. It was recently demonstrated that a simple reverse transcriptase multiplex ligation-dependent probe amplification (RT-MLPA) technique using a 14-transcript panel was concordant with histopathology in 77% of the samples and provided complementary information to the pathologist for a correct diagnosis in most of the remaining non-concordant cases.5 The initiative of the commercially available Banff human organ transplant gene panel studies 770 genes, which opens up further possibilities for multicentre validation studies and the use of molecular diagnostics for future therapeutic decision-making.6
New developments of EMB diagnostics are not limited to the implementation of new molecular techniques; the 21st century has also arrived in basic histology. In the present issue of the European Heart Journal, Peyster et al.7 present an automated computational pathology model (denoted CACHE) for histological grading of cardiac allograft rejection in digital histopathology slides.
An increasing number of pathology labs have implemented a digital workflow where the conventional microscope is replaced with viewing of digital whole-slide images on a computer monitor. This has been enabled by the wide availability of fast and cost-effective whole-slide image scanners that can create high-resolution digital representations of conventional glass tissue slides. In addition to the many ergonomic, patient safety, and workflow advantages, such as remote work that has been of particular significance during the coronavirus disease 2019 (COVID-19) pandemic,8 the availability and use of digital slides offer the possibility of integration of automatic image analysis techniques that aim to supplement or even completely replace diagnostic and prognostic assessments.9
Earlier, these techniques used to rely on classical image analysis and machine learning, and more recently primarily on deep machine learning/artificial intelligence (AI) methods.10 Classical machine learning methods for image analysis require manual engineering of image features that are used as the input to build classification or regression models. In contrast, deep learning models work in an end-to-end fashion, using unprocessed image data directly as the input to the model. Their success lies precisely in this ability to learn relevant hierarchical feature representations of the underlying image content directly from the data. Computational pathology methods based on deep learning have achieved expert-level performance in a variety of cancer histopathology tasks such as detection of breast cancer lymph node metastases and grading of prostate cancer, as well as novel applications such as prediction of genetic mutations from tissue images.11
The CACHE computational pathology model was developed and evaluated using an impressive number of 2476 EMB slides to reproduce the four-grade clinical standard for ACR diagnosis. Interestingly, the model is based on classical machine learning methods using quantitative histological features describing the density and orientation of lymphocytes, myocytes, and stroma. CACHE results were compared with the grades-of-record from the clinical chart and in a subset with the results of blinded re-grading by three independent pathologists. The pathologists that performed re-grading achieved a 60.7% agreement with the grade-of-record, comparable with the pair-wise agreement amongst all human graders of 61.5%. The CACHE model achieved 65.9% agreement with the grade-of-record and a 62.6% agreement with all human graders, which puts it on a par with the interobserver agreement of pathologists. CACHE grading was also found to be insensitive to intercentre variations in tissue processing and digitization, and showed superior sensitivity for high-grade rejection.
These results are very promising and call for taking further steps to push this methodology towards clinical application. Undoubtedly, in the next decade, computer-aided image analysis methods will be further implemented as diagnostic aids for pathologists, and practical approaches such as the CACHE study are instrumental in achieving this. However, as also indicated by the authors, several issues remain open. First of all, there is the question of the ‘gold standard’ used for training of the model. The authors used the ‘grade-of-record’ according to the ISHLT criteria that has high intergrader variability even between expert pathologists.12 It should be stressed that any computational method based on this imperfect standard of visual interpretation in diagnostic pathology might not be able to substantially improve upon the standard itself. An improved gold standard would have been the consensus of multiple experts and the use of additional diagnostic tests, and ideally the inclusion of patient outcome. Nevertheless, these so-called non-inferiority models can still represent a valuable tool, e.g. for helping less experienced pathologists in daily diagnostics or in research studies. Furthermore, the model is limited to ACR. AMR is also an important mechanism of cardiac allograft rejection. When a pathologist has suspicion of an AMR on a routine haematoxylin and eosin (H&E) stain, additional immunostains will be done according to the guidelines of the ISHLT. Until AMR diagnosis is implemented in the model, histological diagnostics by an experienced cardiovascular pathologist will be necessary in routine diagnostics. Next, the CACHE model is built using classical machine learning methods with ‘handcrafted’ image features. The authors consider this an advantage over AI methods that can be difficult to interpret and can potentially include noisy or confounding variables. Still, the use of hand-crafted features does not directly translate to interpretability, and such models might still require visualization tools to properly explain the prediction to pathologists. At the same time, significant progress has been made in the development of deep learning models that are interpretable.13 A comparative study with AI on the same dataset would be an interesting next step. Lastly, there is the question of further validation of this method in order to meet the standard for use in clinical practice. Although the CACHE model was developed and validated in a multicentre setting, only two whole-slide image scanners were used, both approaching end of life. Before clinical translation, it is important to evaluate the generalization performance to different next-generation slide scanners.
The full potential of these algorithms can only be realized in clinical practice with a fully digital diagnostic workflow that few pathology labs around the world have at the moment.14 , 15 This entails scanning of the glass slides as the last step of the routine tissue preparation process, with the digital slides being available upfront before the glass slides reach the desk of the pathologists. Automatic image analysis algorithms can then seamlessly be integrated in the Picture Archiving and Communication System (PACS), and the results from the analysis can be presented during the actual diagnostic process. Scanning individual slides or making digital snapshots with a camera mounted on the microscope for the particular purpose of running automatic image analysis are much less attractive alternatives, but could be used in labs that do not yet have a fully digital diagnostic workflow.
Altogether, CACHE is an important first step in the implementation of computer image analysis in histological cardiac allograft rejection diagnosis, deserving further development and validation. For now, such methods may aid, but certainly not yet replace, pathologists.
Conflict of interest: none declared.
The opinions expressed in this article are not necessarily those of the Editors of the European Heart Journal or of the European Society of Cardiology.
Footnotes
† doi:10.1093/eurheartj/ehab241.
References
