Computational Pathology Assessments of Cardiac Stromal Remodeling: Clinical Correlates and Prognostic Implications in Heart Transplantation

Both overt and indolent inflammatory insults in heart transplantation can accelerate pathologic cardiac remodeling, but there are few tools for monitoring the speed and severity of remodeling over time. To address this need, we developed an automated computational pathology system to measure pathologic remodeling in transplant biopsy samples in a large, retrospective cohort of n=2167 digitized heart transplant biopsy slides. Biopsy images were analyzed to identify the pathologic stromal changes associated with future allograft loss or advanced allograft vasculopathy. Biopsy images were then analyzed to assess which historical allo-inflammatory events drive progression of these pathologic stromal changes over time in serial biopsy samples. The top-5 features of pathologic stromal remodeling most strongly associated with adverse outcomes were also strongly associated with histories of both overt and indolent inflammatory events. Our findings identify previously unappreciated subgroups of higher- and lower-risk transplant patients, and highlight the translational potential of digital pathology analysis.


Introduction
2][3][4] Endomyocardial biopsies (EMB) are routinely obtained, undergoing histologic grading based on the number, extent, and impact of in ltrating immune cells 1,5 .[8][9] While in ltrating immune cells are a primary cause of allograft injury, these cells themselves may not re ect the persistent allograft injury that has occurred.The morphologic correlates of the sustained allograft dysfunction which may result from in ammatory insults are mainly stromal processes affecting the extracellular matrix (ECM), [10][11][12][13] and are subtle and di cult to standardize with conventional histologic assessments.The cardiac stroma consists predominantly of broblasts, which undergo activation and differentiation under in ammatory conditions, leading to ECM remodeling via the deposition of thicker, more disordered collagen bers. 14,15[20] There are no existing tools for measuring the micro-architectural changes that result from in ammatory processes in cardiac allografts.We assert that this represents an important limitation, both in documenting the effects of acute immune-mediated insults such as severe rejection events, and in tracking the longer-term effects of more indolent, sub-clinical processes such as recurrent low-grade rejection or Quilty lesions.An assay capable of detecting the early, microscopic sequelae of pathologic remodeling could identify at-risk patients long before the development of overt, symptomatic, allograft dysfunction.This in turn could provide opportunities for early interventions, either in the form of intensi ed immunosuppression to quash indolent in ammation, therapeutics which have established roles in blunting pathologic remodeling in native hearts or emerging approaches that directly target stromal remodeling. 21 this study, we leverage computational pathology analysis to deeply interrogate the stromal microarchitecture of transplant EMB histology slides.Unlike recent studies utilizing computational pathology to assign International Society for Heart and Lung Transplantation (ISHLT) grades to EMB slides, the present study does not seek to reproduce work already performed by pathologists. 5,22Instead, the focus is on developing a morphologic assay capable of extracting previously unavailable information from EMB histology samples via a rigorous, quantitative, rst-in-eld analysis of allograft stroma.Deploying this novel stromal assessment tool in more than 2000 EMBs, we rst establish the set of stromal 'morphologic biomarkers' most strongly associated with long-term adverse allograft outcomes.We then evaluate how these novel morphologic biomarkers change over time following various allo-immune insults, assessing the impact on stromal remodeling of both overt allograft injury in the form of treated high-grade rejection, and the impact of untreated, indolent in ammation in the form of recurrent, low-grade rejection and Quilty lesions.By quantifying the speci c impact on a tissue-and patient-level of these historical insults, we provide new insights into allograft biology and challenge conventional wisdom on which allo-immune processes merit therapeutic intervention.

Study cohort description, image quality control and color normalization
The study cohort consisted of 2167 hematoxylin and eosin (H&E) stained EMB histology slides obtained from 650 patients treated between 1999 and 2016 at the Hospital of the University of Pennsylvania (UPenn).4][25] Due to our interest in monitoring serial samples and progressive morphologic changes, EMBs from patients who had not received at least 3 EMBs post-transplant were excluded.Due to our interest in the impact of historical -rather than activerejection events on allograft stroma, all study EMBs were con rmed to be free from serious rejection.Speci cally, we excluded EMBs with any acute cellular rejections (ACR) ≥ grade 2R, any antibody-mediated rejections (AMR) with a pathological AMR grade ≥ grade 1, 1,4 and all treated rejection regardless of grade.Histologic and clinical rejection diagnoses (cellular and antibody rejection grades, treated rejection history, and 'Quilty' lesion presence) for all preceding EMB events were compiled for each study EMB, with totals for each diagnosis aggregated.Study EMBs were then assigned to the following 'in ammatory history' subgroups: 1) all patients with a previous high-grade rejection (≥ 2R, ≥pAMR-1), or a treated rejection not otherwise speci ed), 2) the subset of the subgroup 1 patients without recurrent low-grade or Quilty, 3) patients with recurrent low-grade rejection (≥ 3 grade 1R EMBs without any history of high-grade/treated rejection), 4) the subset of the subgroup 3 patients with recurrent low-grade and recurrent Quilty lesions (with ≥ 3 occurrences of each diagnosis) and 5) frequent in ammatory events (any patient with > 5 prior histologic diagnoses of Quilty, ACR ≥ 1R, and/or AMR ≥ pAMR-1), and 6) controls, who met none of the other group inclusion criteria.See Fig. 1 for additional cohort details.Clinical data was collected for all patients contributing an EMB to the cohort, including selected, established or purported donor and recipient risk factors for adverse allograft outcomes as listed in Table 1.Using these data, we identi ed patients with or without "adverse outcomes" de ned as allograft loss (cardiovascular death or re-transplantation) or ≥ grade 2 CAV by 7-years post-transplant.This research complies with the Declaration of Helsinki, and access to archival data and tissue was approved by the University of Ŧ De ned based on number of mismatches in HLA-A, HLA-B, and HLA-DR, as per United Network for Organ Sharing database records.§ De ned as the percentage of patients who contributed a biopsy to the cohort after having a diagnosis of denovo donor speci c antibody with mean uorescent intensity > 1500 All slides were digitized via whole-slide scanning at 40x magni cation.Digitized slides underwent quality control assessments using HistoQC, an open-source, quantitative digital pathology analysis software tool for identifying artifacts and measuring slide quality. 26A total of 265 slides were excluded from the study due to signi cant staining artifacts or damage to the archival slide.Color normalization was applied to each slide to account for slide and batch variations that can negatively impact image segmentation performance. 27romal ber detection and feature analysis Considering that different in ammatory processes act on different areas of cardiac stroma, 2,11,12,14,17,18 three distinct stromal sub-regions were considered for subsequent morphologic feature extraction: 1) endocardial stroma, 2) interstitial stroma, and 3) replacement stroma (Fig. 2b).A total of 210 stromal features were extracted from each EMB slide.Broadly, these features pertain to four distinct categories, as outlined in Fig. 2, and described in more detail in Table 2 and Supplemental Methods.The rst category describes the shape and size of stromal bers: ber length, ber thickness (the total pixel size of the ber divided by the ber length), ber solidity (the pixel ratio of a ber and its smallest convex polygon), and ber perimeter (obtained by the eight-direction chain code).The second feature category characterizes spatial orientation -the degree of local order/disorder -of stromal bers, calculated locally within a 200-pixel neighborhood.The third category describes the accumulation and proliferation of stroma, including spatial density of the bers (the number of stroma bers in a certain pixel area) and the area ratio of cardiomyocytes to interstitial stroma.The fourth category describes the interaction between stromal bers and cardiomyocytes, including assessing the angle and proximity of stromal bers to the surrounding cardiomyocytes.
Table 2 The speci c meaning, quantity, and source of the stromal biomarkers designed in this paper.

2X1X5 = 10
Segmentation method of stromal bers: First, a U-net deep learning architecture was trained to segment cardiomyocytes.The U-net model with an encoding and a decoding component was executed in PyTorch framework on a Titan XGPU running CUDA 7.5 using previously described parameters. 28After obtaining the U-net model's binary masks of the cardiomyocytes, the stroma and the blank areas were morphologically processed by disc-dilation, disc-corrosion, and disc-dilation operations.The binary masks after image morphology operations were then subtracted from the original masks to complete the partitioning of the stromal compartment.Figure 2b illustrates the effect of stroma being divided into three types.A local difference-local binary pattern operator 29 combined with the Otsu algorithm was employed to detect and segment stromal bers -an approach with established success for collagen ber detection in other tissues. 30,31The success of automated stromal ber segmentation was con rmed by experienced pathologists during pipeline development.An illustration of the stromal ber segmentation is depicted in Fig. 3c.Source code for the digital pathology image analysis pipeline is freely available on Github.

Data Analysis and Statistical Methods
To identify morphologic biomarkers of stromal remodeling, we compared the time-dependent response of each extracted stromal feature between the composite Adverse Outcome group (allograft loss or CAV grade ≥ 2 at 7years) and the No Adverse Outcome group.Speci cally, we constructed a mixed-effect regression model which includes variables for group, time, and the interaction term between group and time. 32In this model, the dependent variable was the examined stromal feature, and the independent variables were group and transplant time, with the interaction term formed by group and transplant time considered as a covariate to examine how stromal features respond to transplant time between different groups.By tting this model, we were able to assess for signi cant differences in the stromal feature changes over time-from-transplant between the Adverse Outcome group and the No Adverse Outcome group.The ve features that exhibited the most signi cant changes over time were selected and de ned as the top features of 'pathologic stromal remodeling'.Using these features of pathologic stromal remodeling, we examined the in ammatory history subgroups, comparing stromal feature change-over-time in EMBs from patients in each subgroup vs. Controls.Lastly, in ammatory history subgroups were individually assessed for event rates using the composite Adverse Outcomes data.All statistical analysis was conducted in python math package and Stata v.15.0 (StataCorp LLC). Figure 1 summarizes the experimental work ow from image analysis to data analysis.

Cohort Summary
Summary clinical data for the study cohort is shown in Table  Ŧ CPRA = calculated panel reactive antibody score, a measure of transplant recipient sensitization.§ Positive donor speci c antibody was de ned as an antibody with mean uorescent intensity > 1000.
Prioritizing the morphologic biomarkers of stromal remodeling The top-5 stromal morphologic features associated with the composite Adverse Outcomes are listed in Fig. 3.
Brie y, these stromal features describe the ber density, solidity, and ber orientation of the endocardial stroma, ber thickness of interstitial stroma, and the area-ratio of cardiomyocytes to interstitial stroma.These ve features each exhibit statistically signi cant differences (p < 0.05 for all) in their trends-over-time from transplant between the No Outcome group and the Adverse Outcome group using mixed-effects regression analysis (see Fig. 3b).A more simplistic approach to analyzing and interpreting stromal change based on per-patient averaging features measurements from each EMB a patient contributes to the cohort is shown in Fig. 3a.
Notably, though some signi cant differences between the No Outcome and Adverse Outcome group persist in these averaging-based analyses, much of the nuance is lost with this approach (especially for stromal ber thickness measurements).The feature measurement-over-time slopes presented in Fig. 3b better capture the progressive nature of stromal changes than the averaged feature measurements presented in the bar graphs in Fig. 3a, and represents a more statistically powerful and biologically meaningful method for studying quantitative morphologic data.
Figure 3c provides an intuitive visualization of the differences in stromal features.Examination of non-interstitial stroma in the second and third columns shows that the endocardial stroma from the No Outcome group is more compact, with bers which are shorter, less thick, less parallel, and with more delicate crosslinks.Examination of interstitial stroma in the fourth and fth columns shows that the area, density and length of interstitial stromal bers are signi cantly increased in the Adverse Outcome group, suggesting proliferative interstitial remodeling.
Exploring the effects of previous in ammatory events on stromal remodeling Plots of the change-over-time for each morphologic biomarker of pathologic remodeling in each historical in ammation subgroup vs. controls are illustrated in Fig. 4.There are numerous statistically signi cant differences in the slopes of the feature-change-over-time plots, suggesting that alloimmune events detected on prior biopsies induce measurable stromal changes which can be detected on subsequent biopsies.Notably, signi cant differences in pathologic remodeling were observed not only as a result of long-recognized riskfactors like a history of high-grade/treated rejection, but also in the indolent in ammation subgroups which have experienced recurrent 1R events and/or recurrent Quilty.Additionally, when EMBs which also have a history of recurrent, indolent in ammation (eg.Recurrent 1R or Quilty) are excluded, a history of high-grade/treated rejection alone no longer manifests signi cant differences in most pathologic stromal remodeling features compared to Controls.
Examination of composite outcomes among patients contributing EMBs lends additional context to the in-situ stromal biomarker data.As seen in Table 4, patients contributing EMBs to several of the historical in ammation subgroups have signi cantly higher rates of adverse outcomes compared to patients contributing Control biopsies (5.1% event rate).These include patients contributing historical high-grade/treated rejection biopsies (32.9%, p < .001),patients contributing recurrent low-grade and recurrent Quilty biopsies (12.7%, p = 0.047), and patients contributing recurrent in ammation biopsies (19.5%, p < 0.001).The recurrent low-grade alone group demonstrated an event rate nearly double control patients (9.4% vs. 5.1%), though this was not signi cant (p = 0.16).Once again, a history of high-grade/treated rejection without a concurrent history of recurrent indolent in ammation did not confer increased risk (7.1%, p = 0.67).Taken together, the EMB and patient subgroups analyses suggest that in ammatory insults such as recurrent Quilty lesions and low-level rejection -which are typically not treated in clinical practice -induce measurable changes to the allograft stroma and may be associated with poor patient outcomes.

Discussion
In this manuscript, we describe the development of a computational pathology analysis pipeline designed to comprehensively characterize the stromal architecture of cardiac allografts.Evaluating this pipeline on a large cohort of heart transplant EMBs, we examined the effects of allo-immunity from a novel perspective; focusing on the chronic stromal changes induced by in ammatory insults rather than on the in ammatory cells that induce those changes.Traditional histologic assessments of transplant EMBs such as ISHLT rejection grading focus predominantly on in ltrating in ammatory cells and their immediate effects. 1 Recent computational pathology research in transplant medicine has focused predominantly on reproducing these traditional histologic assessments, 5,22 and as a result, are largely constrained to the same, well-documented, limitations as the ISHLT grading framework.Our approach highlights the value of moving beyond this conventional framework, leveraging digital image analysis to monitor subtle morphologic changes occurring in EMBs over time, then identifying the core set of morphologic changes which portend poor long-term patient outcomes.We assert that future applications of digital pathology would bene t from adopting a similar approach, utilizing longitudinal samples and statistics to correlate progressive morphologic changes with hard clinical endpoints.
From a histopathology perspective, the morphologic features of pathologic remodeling we identi ed provide a detailed view of how the cardiac stroma is changed by different in ammatory insults.Historical in ammatory insults resulted in an increase in interstitial stroma area relative to myocyte area, a nding consistent with the myocyte loss and interstitial brosis that can result from immune-mediated myocardial injury. 5,11,15,18,20In addition, relative to controls, historical in ammatory insults increased stromal ber thickness, solidity, and parallelism (eg.less disordered/branched bers).This may be explained by progressive deposition of type I collagen after an in ammatory injury.Type I and type III collagen are the main structural constituents of the cardiac ECM, with type I collagen manifesting thicker, straighter, and more parallel bers while type III collagen manifests ner, wavier, and more intricately branched bers. 10,17,18It has been shown that rejection and other in ammatory insults can cause the proportion of type I collagen to increase relative to the type III collagen, 11,15,17,18 resulting in a 'stiffer' myocardium.This may explain both the aforementioned stromal features which differentiated EMBs with more/more severe historical in ammatory events from Controls, and may explain the poorer long-term outcomes associated with these features.Lastly, although stroma area was increased, the apparent number/density of bers in the stroma was reduced in EMBs with historical in ammation.Whether this results from different collagen subtypes, from non-brous ECM proliferation, from edema due to indolent in ammation, or from increased stroma 'cellularity' (which contributes to stroma area while increasing the space between individual bers), cannot be de nitively answered from this study.However, each is a potential mechanism worthy of exploration in future research.
The experiments reported in this manuscript yielded several ndings of translational value.First, identifying the progressive stromal changes which are most strongly correlated with future adverse outcomes creates opportunities for intervention, either through augmented immunosuppression, through the use of traditional heart failure therapeutics with 'reverse remodeling' capability, or by application of new treatments which directly target stromal remodeling. 21Whether the speci c biomarkers of pathologic remodeling uncovered in this experiment can be used to monitor treatment effects after a therapeutic intervention remains unknown, but is an additional potential application for the novel stromal biomarkers reported in this paper.Moreover, the nding that recurrent low-grade in ammatory processes are linked to adverse long-term outcomes is signi cant and worthy of further discussion.
It is common practice for transplant clinicians to monitor, but not treat, low-grade ACR events and Quilty lesions, only implementing acute or chronic therapeutic interventions for cases involving clinical evidence of allograft dysfunction.In fact, current ISHLT guidelines generally discourage treatment of low-grade ACR events. 33On the other hand, most episodes of 'high-grade' rejection as de ned in this manuscript (either ACR ≥ 2R or pAMR > 0) undergo either acute treatment or alterations of chronic immunosuppression, largely in accordance with existing guidelines. 33,34In the present study, our results show that recurrent, indolent in ammatory processes like lowgrade ACR and Quilty lesions are associated with signi cant, pathologic changes in the cardiac stroma, and that this leads to a higher incidence of adverse allograft outcomes.In addition, our results showed that isolated episodes of high-grade rejection in patients without a history of recurrent 1R or Quilty do not appear to induce signi cant long-term pathologic changes in the cardiac stroma.In the context of current practice patterns and guidelines, these ndings suggest that clinicians may be under-valuing the importance of chronic 'mild' alloimmune responses, and may be -in some cases -over-valuing the impact of isolated, high-grade histologic rejection.
Prior research has shown correlations between Quilty lesions and adverse outcomes, though the available literature has con icting ndings. 1,2,20,35The impact of recurrent low-grade ACR on transplant outcomes has not been studied as frequently, though recent research does suggest that a history of higher 'average' rejection grades (even in the absence of high-grade events) is associated with a higher incidence of early CAV. 20While in this manuscript our ndings generally support a connection between recurrent, untreated, indolent in ammation and adverse events, it is clear that not all patients with a history of Quilty and/or low-grade ACR necessarily suffer poor outcomes.Moreover, due the retrospective nature of this research, there is no way to assess the potential risks or bene ts that might arise from altering immunosuppression based on a history of recurrent indolent in ammation.Nevertheless, given the strong correlation between pathologic remodeling features and poor outcomes in these patients, it is worth considering a potential clinical role for our stromal biomarkers.EMB samples are already obtained as part of routine care, and digital pathology analysis pipelines can be quickly and remotely accessed through cloud-based systems.Thus, while future clinical investigations are clearly needed, protocols which incorporate predictive morphologic biomarkers into immunosuppression management and CAV screening decisions might prove to be feasible and valuable.
As the eld gradually pivots towards rejection surveillance paradigms which utilize more 'liquid biopsy' serologic assays and fewer EMBs, [6][7][8][9] we assert that it will become increasingly important to rely on digital pathology biomarkers like those in this manuscript.If patients are to receive only 3-4 EMBs during their post-transplant course, then it is critical to extract maximum information from each of these events.The fewer EMBs performed, the lower the likelihood of identifying patients who are experiencing poor-outcome-associated recurrent 1R and Quilty lesions.Thus, it will be necessary to rely on surrogates for these recurrent histologic diagnoses, such as the biomarkers of pathologic stromal remodeling which we identi ed in this manuscript and which have clear associations with adverse outcomes.Future rejection surveillance protocols could therefore rely primarily on serologic testing, with EMBs performed at a few, widely spaced intervals to enable monitoring of subtle, serial changes which help identify at-risk populations.Compared to traditional, biopsy-heavy approaches relying on conventional histologic grading, such a hybrid approach could maximize personalization while still minimizing invasive testing.
As with all research, this study has limitations.Although the cohort comprised over 2000 biopsies, this was a single center study, and there were relatively few patients in the interesting 'previous high-grade rejection without recurrent low-grade or Quilty' subgroup.Additionally, while we utilized all available histologic diagnoses associated with study EMBs, additional, unmeasured, allo-immune processes could have confounded our ndings.Due to limited application of immunostaining at our center on routine screening EMBs, our historical cohort precluded a complete and de nitive assessment of AMR for in many cases.While we can con dently evaluate whether histologic criteria for AMR are met (i.e.pAMR(h+)), historical assessments of pAMR-(i+)) are limited to those EMBs which underwent clinical immunostaining at the time of EMB.Nevertheless, without pAMR(h+), without concurrent positive donor speci c antibody testing, and without clinical evidence of rejection or provider decision to treat for rejection, we are con dent that major rejection events were not mislabeled as a result of our center's practice of intermittent/for-cause use of immune-staining on routine surveillance EMBs.
Another unavoidable limitation of this study is reliance on pathology diagnostic records for assigning 'in ammatory history' case labels.It is well known that there is signi cant inter-pathologist variability in the application of ISHLT grades to transplant EMBs. 5,36,37Thus, study labels like 'previous high-grade rejection' or 'recurrent low-grade rejection' (without a history of high-grade rejection) are not de nitively accurate.The fact that different pathologists would likely grade historical EMBs differently means that there is inevitable overlap between some study subgroups.It should be noted that while this limitation may affect subgroup comparisons, it has no impact on the correlations between speci c patterns of stromal remodeling and patient-level clinical outcomes -a fact which further highlights the need for grounding computational pathology research in de nitive clinical endpoints rather than imperfect histologic reference standards.
In conclusion, this study represents a novel and important application of computational pathology analysis within heart transplant medicine.Focusing on the allograft tissue itself rather than on the in ltrating immune cells, the stromal morphologic biomarkers described in this manuscript demonstrate the ability to quantify the effects of various historical in ammatory insults, uncovering new information about how different histories may predispose patients to adverse clinical outcomes.

1 )
: pixels in ber orientation 2) Thickness: average pixels in perpendicular to ber orientation 3) Size: total number of pixels 4) Extent: ratio of pixels in the region to pixels in the total bounding box 6) Eccentricity: ratio of the focal distance over the major axis length 7) Solidity: ratio of pixels in the region to pixels of the convex hull image Angle between Interstitial ber and cardiomyocyte surface, or between endocardial ber and endocardial border.

Figure 4 Time
Figure 4 1. Overall, 71/650 patients (10.9%) experienced the composite Adverse Outcome.As expected, there were numerous clinical risk factors associated with the adverse outcomes of allograft loss or CAV grade ≥ 2 at 7-years, though Pearson correlation coe cients were generally modest (Table3).The correlation coe cients of six clinical risk factors with composite outcomes exceeded 0.1, revealing a weak correlation.These six risk factors are re-transplantation, BMI at 1-year, Diabetes, CKD, Donor age, and Donor hypertension.

Table 3
Correlation coe cients for clinical variables with composite adverse outcomes * HLA = Human leukocyte antigen.Mismatch de ned by mismatch between donor and recipient as captured in the United Network for Organ Sharing records, which speci cally captures mismatches at HLA loci A, B, and Dr.

Table 4
Adverse outcome event rate for patients contributing biopsies to each historical in ammation subgroup