The prognostic value of tumor-stroma ratio and a newly developed computer-aided quantitative analysis of routine H&E slides in high-grade serous ovarian cancer

Introduction: Tumor-stroma ratio (TSR) is prognostic in multiple cancers, while its role in high-grade serous ovarian cancer (HGSOC) remains unclear. Despite the prognostic insight gained from genetic profiles and tumor-infiltrating lymphocytes (TILs), the prognostic use of histology slides remains limited, while it enables the identification of tumor characteristics via computational pathology reducing scoring time and costs. To address this, this study aimed to assess TSR’s prognostic role in HGSOC and its association with TILs. We additionally developed an algorithm, Ovarian-TSR (OTSR), using deep learning for TSR scoring, comparing it to manual scoring. Methods: 340 patients with advanced-stage who underwent primary debulking surgery (PDS) or neo-adjuvant chemotherapy (NACT) with interval debulking (IDS). TSR was assessed in both the most invasive (MI) and whole tumor (WT) regions through manual scoring by pathologists and quantification using OTSR. Patients were categorized as stroma-rich (≥ 50% stroma) or stroma-poor (< 50%). TILs were evaluated via immunohistochemical staining. Results: In PDS, stroma-rich tumors were significantly associated with a more frequent papillary growth pattern (60% vs 34%), while In NACT stroma-rich tumors had a lower Tumor Regression Grading (TRG 4&5, 21% vs 57%) and increased pleural metastasis (25% vs 16%). Stroma-rich patients had significantly shorter overall and progression-free survival compared to stroma-poor (31 versus 45 months; P < 0.0001, and 15 versus 17 months; P = 0.0008, respectively). Combining stromal percentage and TILs led to three distinct survival groups with good (stroma-poor, high TIL), medium (stroma-rich, high TIL, or; stroma-poor, Low TIL), and poor(stroma-rich, low TIL) survival. These survival groups remained significant in CD8 and CD103 in multivariable analysis (Hazard ratio (HR) = 1.42, 95% Confidence-interval (CI) = 1.02–1.99; HR = 1.49, 95% CI = 1.01–2.18, and HR = 1.48, 95% CI = 1.05–2.08; HR = 2.24, 95% CI = 1.55–3.23, respectively). OTSR was able to recapitulate these results and demonstrated high concordance with expert pathologists (correlation = 0.83). Conclusions: TSR is an independent prognostic factor for survival assessment in HGSOC. Stroma-rich tumors have a worse prognosis and, in the case of NACT, a higher likelihood of pleural metastasis. OTSR provides a cost and time-efficient way of determining TSR with high reproducibility and reduced inter-observer variability.


Introduction
Ovarian cancer is the most lethal gynecological malignancy in developed countries.(1) In most cases, the disease is diagnosed at an advanced stage and relapse is common, resulting in 8-year survival of less than 20%.(2) It remains a challenge to identify patients in whom treatment is likely to fail, and reliable biomarkers for predicting treatment failure and survival are urgently needed.In an attempt to address this need, the in uence of the tumor microenvironment (TME) and molecular pro les on survival has become a topic of interest, particularly in high-grade serous ovarian carcinoma (HGSOC).(3)(4)(5) While studies on the TME and molecular pro ling have changed the way we think about ovarian cancer heterogeneity, (3)(4)(5) the translation of the knowledge of the found tumor subtypes into the clinical setting is often limited by the cost and scale of molecular pro ling, even more so in developing countries.Besides cost-associated issues, not all samples meet the needed quality or quantity required for such tests.Thus, cost-effective and generally applicable alternatives could accelerate the translation of new research discoveries and the development of quantitative biomarkers for ovarian cancer.
It has often been described that the development and progression of ovarian cancer are in uenced by the interaction between tumor cells and the microenvironment, mainly conferred by the tumor stroma.(6)The stroma is composed of the extracellular matrix and connective-tissue cells such as broblasts and mesenchymal stromal cells.(7)A higher stromal percentage as re ected in the Tumor-Stroma Ratio (TSR), has been associated with tumor growth, metastasis, chemoresistance, and recurrence in multiple epithelial cancers.Tumor stroma is hypothesized to confer its tumor-promoting effect via different mechanisms such as extracellular matrix remodeling, suppression of immune cells, and alteration of stromal regulatory pathways.(8)(9)(10)(11)(12)(13) Pro ling of ovarian cancer based on TSR with the use of Hematoxylin and eosin (H&E) stained slides has been reported infrequently,(14-16) despite its cost-e ciency, DNA independence, and depicted prognostic role in other cancer types.Additionally, studies that have investigated the in uence of manually determined TSR have focused on the most invasive (MI) tumor region, (14,15) while the few studies based on automated image analysis have focused on the whole tumor (WT), limiting comparability between both scoring techniques.(16,17) Advances in the infrastructure of digital pathology, improvement of deep learning models, and the existence of national databases result in the opportunity to further develop automated image analysis methods.
The ability to objectively identify stromal components via computational pathology could enable computer-aided diagnosis(18), complement molecular analysis in a cost-e cient manner, and uncover new therapeutic targets.It also reduces inter-observer variability allowing a more robust categorization of patients.Therefore, in the present study, we aimed to determine the clinical implication of TSR in HGSOC and to develop an accurate image analysis classi er (Ovarian-TSR; OTSR).

Materials and Methods
Patient selection.360 patients with advanced stage (FIGO stage IIb-IV) HGSOC treated with a combination of cytoreductive surgery and chemotherapy from the Netherlands Cancer Institute -Antoni van Leeuwenhoek Hospital (NKI-AVL), Maastricht University Medical Centre (MUMC) and Amsterdam University Medical Centre (AUMC), between January 2008 and December 2015, were included.Of these, 141 patients received primary debulking (PDS) and adjuvant chemotherapy (NKI-AVL n = 52, MUMC n = 30, AUMC n = 59), and the remaining 219 patients neoadjuvant chemotherapy (NACT) followed by interval debulking surgery (IDS) (all NKI-AVL) (Supplementary gure S1A).(19) Clinical data was extracted from the Netherlands Cancer Registry (NCR), a nationwide registry covering all primary malignancies in the Netherlands, and histopathological data from the Dutch Pathology Registry (PALGA).For the present study, approval of the institutional review boards of the NCR (reference number; K22.367), PALGA (reference number; 2016-82), and NKI-AVL (reference number; CFMPB297), was obtained.Ethical approval was waived according to Dutch legislation.The following parameters were extracted from the NCR; performance status, germline BRCA status, treatment sequence (NACT, NACT-IDS, or PDS), surgery outcome (complete with no visible disease; optimal with ≤ 1 cm residue or sub-optimal with > 1 cm residue), distant metastasis (grouped based on localization as pleural malignant effusion, parenchymal, extra-abdominal lymph node, and "other" if there were metastases in multiple categories), and progression status.Progression of disease was de ned in the case of symptoms combined with increased serum CA-125 levels, radiological signs of progression, or histological or cytological con rmation of recurrent disease.Vital status and date of death were obtained via linkage with the municipal population registration.

Hematoxylin and eosin (H&E) image processing
Representative samples of primary ovarian cancer from formalin-xed, para n-embedded (FFPE) tissue blocks of the 360 patients were obtained from PALGA (Supplementary gure S1B) and were either pretreated or chemo-naïve depending on therapy sequence.
All cases underwent central pathological review by three dedicated pathologists (KVdV, HH, JS).Tumor Regression Grade, similar to the Mandard score, of pretreated slides was determined.(20) 4 µm thick tissue sections were taken to perform H&E staining.H&E slides were scanned using an Aperio scanner (Aperio, San Diego, USA) at 40 x yielding digital images with a resolution of 0.25 µm/pixel.All acquired images were anonymized for research purposes.

Manual TSR scores
Each of the H&E-stained whole slide images (WSI) was evaluated by two pathologists (KVdV, JS), who specialized in gynecologic oncology, while being blinded for patient characteristics and outcome.Two TSR scores were determined; the TSR of the MI area of the tumor and the TSR of the WT area.Slides were scored through the online scoring platform Slidescore (https://www.slidescore.com).Both pathologists selected the MI of each slide with the use of a 4x objective.With a 10x objective, the pathologists chose a part of the sample containing both cancer and stromal cells.Cancer cells had to be present at all borders of the image eld.The percentage of cancer and stromal cells were scored while visually excluding necrosis, adipose tissue, psammoma bodies, and vessels.The WT area was determined and scored similarly.The WT area of each slide was selected with the use of a 4x objective and the percentage of cancer cells and stromal in ltrate was scored.The stromal density was evaluated per tenfold percentage (10%, 20%, 30%, etc.).In case of a lack of concordance between both pathologists' scores, a third pathologist (HMH) was consulted to reach a consensus.Patients were further strati ed in stroma-rich (stroma ≥ 50%) or a stromapoor (stroma < 50%), consistent with previous literature.(15) Computational scoring To measure TSR in an unbiased and fully automated way, we developed a deep learning based computational pipeline termed OTSR.OTSR scores were computed both for WT and MI in three steps: 1) segmentation of tissue into relevant tissue types, 2) estimation of the region of measurement (for WT and MI, separately), and nally, 3) measuring the stromal density based on tissue abundance within the region of measurement (Fig. 1).
In the rst step, a deep learning model was employed to classify WSI patches into tumor, stroma, and background (necrosis, adipose tissue, psammoma bodies, and vessels).The custom convolutional neural network was speci cally designed to provide tissue segmentation at high resolutions while taking the surrounding tissue context into account (Supplementary Figure S2).WSI tissue segmentation was subsequently achieved by classifying individual WSI image patches extracted at 20x magni cation.
Image patches were extracted with overlap to achieve high-resolution tissue segmentation.
In the second step, the exact region of measurement was estimated for WT and MI.To measure TSR in the MI region, we rst de ned the MI region computationally as the region within a 1.6 mm radius -corresponding to a 10x microscopy eld of view (FoV) -with the highest tumor percentage, where the tumor percentage is measured excluding the background.To exclude regions beyond the tissue border, FOV was only considered if it was composed of more than 2/3 of non-background tissue.To measure TSR for WT, the region of measurement was based on the tissue of which the local tumor density was higher than a predetermined threshold.For smaller tumors, the threshold was dynamically altered based on the tumor size, leading to a smaller stroma margin around the tumor tissue.The threshold was chosen empirically to correspond to reference tumor bed annotations indicated by the pathologists while conducting the manual TSR measurements through Slidescore.
In the nal third step, TSR was measured as the stromal density within the region of measurement, excluding the background.
Similar to the manual scoring, patients were strati ed in stroma-rich (stroma ≥ 50%) or stroma-poor (stroma < 50%).Additionally, to explore the optimal cutoff for TSR we determined the TSR cutoff with the highest discriminative power for OS in patients who were treated with PDS, for MI and WT.Supplementary text 1 presents a more detailed description of the OTSR pipeline.
Statistical analyses were performed in STATA/SE (version 14.1, STATA CORP, College Station, TX, USA).A p-value < 0.05 was considered statistically signi cant.Clinical associations with TSR ratios and patient characteristics were assessed with Chi-square, or Fisher's exact test if > 20% of expected cell counts are less than 5, for categorical variables, one-way ANOVA for normally distributed continuous variables, and Kruskal-Wallis for non-normally distributed continuous variables.Kaplan-Meier survival estimates with the corresponding log-rank test and univariable and multivariable Cox regression analyses were used to assess the effect of TSR on progression-free survival (PFS) and overall survival (OS).Those found signi cant in univariable analyses with a p < 0.10, were included in the multivariable regression analyses and assessed using backward selection.PFS was calculated as the time between the date of diagnosis of the primary tumor and the date of recurrence.OS was calculated as the interval between the date of diagnosis of the primary tumor and the date of death.Patients without recurrence and/or alive at the date of the last check of the municipal population register (31 January 2022) were right censored for the respective analysis.

Results
Patient characteristics.Tissue samples of 360 patients with FIGO stage III-IV ovarian cancer were identi ed. and collected (Supplementary Figure S1A), of which 344 samples were available for WSI analysis.Four samples contained too little tumor and were excluded from analysis resulting in 340 samples (Supplementary Figure S1B).15 MI cases and 17 WT cases, respectively, were excluded from computational scoring due to too little tissue or poor image quality.After excluding the test cases, 303 computationally scored patients for MI and 301 for WT were available for analysis (Supplementary Figure S1B).Most FFPE blocks were derived from the ovary/tuba (79%).All clinicopathologic characteristics of patients are summarized in Table 1.186 (55%) patients were 65 years or older, and 314 (92%) patients presented with FIGO stage III or higher, at diagnosis, with pleural malignant effusion being the most common distant metastasis.141 (42%) patients underwent PDS and 199 (58%) received NACT-IDS.
Quantifying tumor stroma composition with computational and manual image analysis.
All slides were scored by two independent pathologists depicting a signi cant positive correlation between both scores (MI; Pearson R = 0.93, and Interclass Correlation Coe cient (ICC) = 0.93) (Supplementary Figure S3).112 (33%) patients had a stromarich MI region compared to 175 (52%) WT (Table 1).OTSR resulted in 79 (26%) patients with a stroma-rich MI region compared to 150 (50%) WT.A signi cant positive correlation was seen between manual and OTSR scores (MI; Pearson R = 0.83 and Interclass Correlation Coe cient (ICC) = 0.81) suggesting a high correlation between both methods (Supplementary Figure S4).In patients treated with PDS the optimal computational cutoff, stroma-rich versus stroma-poor, was 18% stroma in MI and 28% stroma in WT resulting in 63% and 88% of patients with a stroma-rich MI and WT region, respectively (data not shown).
Association between TSR and Survival.

Independent prognostic value of TSR
To evaluate whether TSR is an independent prognostic biomarker, we performed multivariable analysis correcting for parameters that were found signi cant in univariable analysis (Supplementary Table S2).Known prognostic factors for ovarian cancer as the completeness of debulking, age, FIGO stage, therapy sequence, and residual status were signi cantly associated with OS.Notably, the localization from where the H&E slide originated, such as the ovary or the omentum, did not in uence OS.When combining both treatment types, TSR remained prognostic in a multivariable analysis for OS, but only statistically signi cant in the case of OTSR when employing the optimal cutoff (MI; HR = 1.39, 95%CI = 1.04-1.84)(Table3).When subgrouping for treatment type, TSR remained signi cantly associated with OS in multivariable analysis in the case of MI (HR = 1.83, 95%CI = 1.07-3.13)and WT (HR = 1.83, 95%CI = 1.18-2.82)(Supplementary Table S1) in manually scored cases of patients treated with PDS.OTSR did not show statistical signi cance in patients treated with PDS, most likely as a result of low patient numbers.In NACT, TSR was not signi cantly associated with OS.PFS also did not remain signi cantly associated with TSR in multivariable analysis (Table 3).

Lymphocyte and stromal cell ratio as a joint prognostic classi er.
Given the known prognostic value of TILs (19) and the previously demonstrated prognostic value of TSR, we investigated if both parameters could de ne a better prognostic classi er when employed in symphony.As suggested by chi-square analysis (Table 2), there was a negative correlation between TSR and TILs (CD68: Spearman correlation coe cient − 0.17, p = 0.0013; CD103: Spearman correlation coe cient − 0.19, p = 0.0003)(data not shown).We selected patients with the highest quartile of TIL densities as the TIL high group.As shown in a previous study, patients with high TIL densities show a signi cantly better OS compared with the rest of the patients, with a 10-year OS probability ranging from 15%-19% (CD8 and CD103, respectively) in high TILs versus 8-9% (CD8 and CD103, respectively) in low TILs.(19) In case of TSR, the 10-year OS probability ranged from 5%-4% in case of stroma-rich to 13%-17% in case of stroma-poor (MI and WT, respectively, Fig. 2A, Fig. 2D.)To assess the possible signi cance of a joint classi er, patients were assigned into three groups; low risk (stroma-poor, high TIL), medium risk (stromarich, high TIL or stroma-poor, low TIL) or high risk (stroma-rich, Low TIL).We observed that patients with high CD8 + densities and stroma-poor (low risk) had a 10-year OS probability of 16%, compared to 11% (medium risk) and 3% (high risk) (HR = 1.35,CI = 0.97-1.87;and HR = 2.15, CI = 1.49-3.11,respectively, Supplementary Table S3, Fig. 4A).The same results were seen in the case of CD103 (10-year OS; 19% versus 11% and 4%), CD20 (10-year OS; 24% versus 12% and 3%), and CD68 (10-year OS; 14% versus 12% and 3%), all in favor of stroma-poor and high TILs (Fig. 4B-D, Supplementary Table S3).These results remained signi cant in multivariable analysis for CD8 and CD103 combined with TSR (Supplementary Table S3).

Discussion
In this study, we presented the prognostic value of TSR in advanced-stage HGSOC both in the most invasive tumor region as well as in the whole tumor.We also identi ed a difference in the prognostic value of TSR in patients treated with PDS followed by chemotherapy versus patients treated with NACT-IDS.We additionally presented an algorithm, OSTR, which computationally quanti es TSR in both regions using H&E-stained slides.Furthermore, we demonstrated a joint prognostic classi er combining TSR and distinct immune cell densities.
There is increasing evidence supporting the importance of the TME in ovarian cancer progression.(3)(4)(5) However, most studies are performed using costly methods such as molecular pro ling.We demonstrate a cost-effective way to study the TME and identify high-risk patients with routinely generated H&E slides.The present study focused on both the most invasive region and the whole tumor, thus differing from studies focusing on only one of both areas.
The key nding of our study is that the stroma-rich feature is signi cantly associated with a poor prognosis in HGSOC, in both the MI region as well as the WT, scored either manually or computationally.TSR can effectively separate patients into two groups, with either a favorable or unfavorable prognosis, independent of and complementary to known prognostic variables.This observation is consistent with the study by Chen et al., in which patients with stroma-rich in the MI region depicted a worse OS and PFS in epithelial ovarian cancer.(15) Furthermore, the results from our automated image analyses are consistent with Lan et al. and Jiang et al., who described an automated image analysis of TSR, using the WT, in which higher percentages of stromal cells were associated with worse survival.(16,17) In addition to the previous studies, we indicated that TSR is a prognostic parameter in both the MI and WT region and that both regions can be computationally scored generating similar results to images scored by pathologists.We also depicted that the TSR is not a prognostic parameter in patients treated with NACT, which could be a result of stromal recomposition.
In daily practice, histological characteristics, such as TSR, are manually determined by individual pathologists.Manual examination of H&E stained slides is in uenced by inter-and intra-observer variability (22), and relies on the pathologistsé xperience and expertise.An unbiased quanti cation is often unrealistic which could result in heterogenous results and in uence reproducibility.The use of computational scoring methods, such as OTSR, eliminates inter-and intra-observer variability.OTSR also eliminates the time that the pathologist needs to dedicate to scoring the H&E slides and is therefore cost-effective.Lastly, OTSR enables the use of an optimal threshold, resulting in a more nuanced and reproducible biomarker allowing for more speci c patient strati cation.
Stromal density has been associated with TIL density, with higher stromal percentages being associated with lower lymphocytic in ltration.(16,23) T-cell in ltration is a well-described prognostic marker in ovarian cancer, in which higher amounts of TILs are associated with favorable survival.(4,24) Our study con rmed the association between TIL and stromal densities.In contrast to previous reports, we depicted distinct TILs, CD103 + and CD68+, that were signi cantly lower in the case of stroma-rich tumors.
Interestingly, in patients treated with NACT, CD20 + densities were positively correlated with TSR, which could be a re ection of the known increase in TILs post-NACT, (25) or a TIL-speci c relationship with TSR.We also demonstrated that immune cell in ltration and TSR could co-de ne a prognostic classi er, in agreement with previous reports.(16)Again, unlike previous studies,(16) we depicted distinct TILs, in this case, CD8 + and CD103+, that were independent prognostic markers when combined with TSR, in which stroma-rich and low TILs had a signi cantly worse OS compared to stroma-poor and high TILs.Taken together, our data suggest that the H&E-based TSR could be intrinsic and clinically relevant in advanced-stage ovarian cancer and could identify distinct morphological subtypes in HGSOC.Nonetheless, more studies are needed to clarify the associations between speci c TILs and TSR.
The strength of the present study is that we integrated detailed pathological and computational scoring of the TME with immunological and clinical data from a large cohort of patients with HGSOC.Additionally, we predicted TSR in both the MI and WT and not only on a whole slide level as previous reports suggest that tissue regions contribute to the prognostic signi cance of TSR.All slides were centrally reviewed by two independent dedicated pathologists in gynecologic oncology.Lastly, a long follow-up was achieved and clinical data were complete.Limitations of the study include its historic nature, heterogeneous treatment regimen, sample size, and lack of a validation set.Furthermore, our study included samples from three separate institutions.H&E imagingbased digital pathology studies may be affected by para n block preservation protocols, which could lead to model differences.Nonetheless, our study showed no association between TSR scores and the institution from which the sample was derived demonstrating the robustness of the computational model.Although the prognostic value of computational and pathological scored TSR was similar, pathological expertise was essential for developing the initial OTSR.Future studies are needed to identify the in uence of stromal subtypes on cancer progression which could be key to new therapeutic targets.
As stated, cancer development is in uenced by the interaction between the tumor and the microenvironment.(6)As a result, the TME has become an attractive target for new therapeutic targets.(26)Our study depicted that TSR, as a part of the TME, is a cost-effective and easy-to-determine prognostic marker based on H&E-stained tumor slides that are generated during routine pathological examination.Assessment of TSR could identify patients with signi cantly worse prognoses who could bene t from novel agents targeting the stromal fraction of the tumor.

Table 1 .
Contributions: L.v.W. and C.W. performed study conceptualization and design, gathering of resources, data curation, statistical analysis, interpretation of data, and drafting of the manuscript; J.S. contributed to the methodology; K.K.V.d.V., A.R.J, and H.M.H. contributed to the methodology, study conceptualization, interpretation of data, reviewing, and editing; S.R., R.F.K., M.A.v.d.A., and G.S.S. contributed to study conceptualization, interpretation of data, reviewing, and editing.All authors have read and agreed to the published version of the manuscript.Institutional Review Board Statement: For the present study, approval of the institutional review boards of the NCR [K22.367],PALGA [2016-82], and NKI-AVL [CFMPB297], was obtained.Ethical approval was waived according to Dutch legislation.Informed Consent Statement: Patient consent was waived according to Dutch legislation.Patient and Tumor Characteristics

Table 2 .
Tumor and patient characteristics by most invasive TSR and treatment type *Based on most invasive TSR scored by pathologists.Abbreviations TSR: tumor stroma ratio; NACT: neoadjuvant chemotherapy; TMA: Tumor micro array; P: P-value

Table 3 .
Survival analysis, Tumor stroma ratio *Data was adjusted for age, FIGO stage, therapy sequence and outcome of surgery