Background
Ovarian cancer is the leading cause of death in patients with gynecologic malignancies. In 2012, there were approximately 238,700 new cases of ovarian cancer and 151,900 deaths in the world [
1]. Most women present with advanced disease at diagnosis. Despite initial treatment with primary debulking surgery followed by first-line taxane and platinum-based chemotherapy, about 75% of patients will eventually relapse [
2]. Disease that relapses ≥6 months after completion of platinum-based chemotherapy is categorized as platinum-sensitive recurrence [
3], while disease that relapses < 6 months after platinum-based chemotherapy are defined as platinum-resistant recurrence [
4]. Some patients initially present with platinum-sensitive recurrent diseases, however, nearly all patients with platinum-sensitive disease ultimately develop platinum resistance, which results in death [
5]. Therefore, platinum resistance is a major determinant of poor outcome in ovarian cancer. Predicting the risk of platinum resistance will provide basis of improved therapeutic efficacy in ovarian cancer.
In the past 20 years, the investigations of chemotherapy resistance were mostly restricted to cancer cell intrinsic events, and the contributory effects of different types of cells coexisting within the tumor were ignored [
6]. However, a large body of evidence has shown the crosstalk between cancer cells as well as stromal cells in tumor microenvironment enable tumor growth and progression [
7‐
11]. Moreover, microenvironment niches can promote chemoresistance within a tumor [
12‐
15]. The abundance and characteristic of stromal cells and immune cells in tumor microenvironment have been demonstrated to lead to chemotherapy resistance in multiple cancers [
7,
15‐
19].
Recently, computer vision techniques have been applied to tumor pathologic specimen and enable automated identification and classification of various cell types in tumor microenvironment [
20‐
23]. These methods make rapid cell mapping in tumor pathologic specimen and quantifying cellular composition across tumor microenvironment. In one of our previous studies, we developed an automated image analysis system for identifying and quantifying heterogeneous cell types in ovarian cancer histological sections [
24]. One of the findings was that high amount of stromal cells in ovarian cancer was significantly associated with poor overall survival.
Therefore, in this paper, we aimed to (i) use automated image analysis of histological sections to identify the association between microenvironmental composition and platinum-resistant recurrence in ovarian cancer, (ii) develop a novel model to predict the risk of platinum-resistant recurrence.
Methods
Patient selection
Ninety-one patients with ovarian cancer containing data of automated image analysis for hematoxylin & eosin (H&E) stained histological sections which were used in our previous study were initially reviewed [
24]. All patients were stage III-IV disease and received primary debulking surgery in our institute.
In these 91 patients, 71 patients had recurrent disease and 20 patients did not. Only the 71 patients with recurrence were enrolled in this study. Clinical data regarding patient demographics, surgical records, pathologic characteristics, platinum-free interval, serum CA125 level (normal level defined as ≤35 U/ml), treatment regimen and follow-up were extracted from the hospital records. Patients with disease that relapses < 6 months after completion of a platinum-based chemotherapy were defined as platinum-resistant relapse. Informed consent was obtained from all patients. Ethical approval were obtained by the institutional review board of Sun Yat-sen University Cancer Center. All methods were carried out in accordance with the approved guidelines of our institute.
Histology image analysis and microenvironmental composition quantification
Automated image analysis of H&E-stained whole-tumor sections was described in our previous study [
24]. In brief, H&E histological sections of ovarian tumors were generated, digitalized and analyzed using the R package CRImage. A feature vector of 97 morphological and textural measurements derived from hematoxylin positive nuclei were all subjected to a support vector machine (SVM) for classification to identify single cells including cancer cell, stromal cell and lymphocyte. The image analysis system was trained using a small set of training data (900 cells provided by an expert pathologist) using SVM. Experiments were performed to validate the system. Then, cancer cells, stromal cells and lymphocytes were quantified in each tumor section. And the cell ratio for each sample was obtained, which was used to build predictive models in this investigation.
Pathological scores
Pathological scores including percentage of cancer cells and stromal cells for each H&E histological section were given by a pathologist. Lymphocytic infiltration was assessed as low, median, or high infiltration.
Statistical methods
Difference in platinum-resistant recurrence between different groups was evaluated using Chi-square test.
The receiver operating characteristic (ROC) curve was used to evaluate the performance of predictive models and was created by plotting the true positive rate (TPR) against the false positive rate (FPR) at various threshold settings. The area under the curve (AUC) provided a value description for the performance of the ROC curve. The AUC score would be between 0 and 1. The higher the value of AUC, the better the model was [
25].
$$ AUC\kern0.5em =\kern0.5em \frac{\Sigma_{i\kern0.5em =\kern0.5em 1}^{n_0}\left({r}_i\kern0.5em -\kern0.5em i\right)}{n_0\times \kern0.5em {n}_1}\kern0.5em =\kern0.5em \frac{\Sigma_{i\kern0.5em =\kern0.5em 1}^{n_0}\kern0.5em {r}_i\kern0.5em -\kern0.5em \frac{n_0\kern0.5em \times \left({n}_0\kern0.5em +\kern0.5em 1\right)}{2}}{n_0\times \kern0.5em {n}_1}. $$
where ri was the rank of the ith positive participant in the ranked list, and n0 and n1 were the numbers of the positive and negative participants.
The predictive value of models including a single parameter for platinum-resistant recurrence was determined using ROC curves. For combined parameters, the predictive models were constructed according to the following procedures. First we fitted multivariate logistic regression model using the characteristics which univariate logistic regression analysis suggested as significant and the status of platinum-resistant recurrence as the dependent variable. Then we calculated the probability P values of each participant using these logistic regression models. Third, we used the calculated P value as a predictor to build the ROC curves. The AUC between different models was compared using Z test by SAS 9.3 software. And the other data analysis were performed by SPSS Statistics 16.0 software (SPSS Inc., Chicago, IL, USA). All the test were two sided and a p value < 0.05 was considered as significant. The key raw data have been recorded at Sun Yat-sen University Cancer Center for future reference (number RDDA2017000306).
Discussion
There is increasing evidence that heterogeneous components in tumor microenvironment collectively promote tumor chemoresistance [
7,
11,
14]. However, most of studies investigated the association between microenvironmental components and chemoresistance by microarrays [
26], sequencing technology [
27], Polymerase Chain Reaction (PCR) [
28] and immunohistochemistry [
29]. Computational analysis of pathological images with quantitative and objective nature provides a new opportunity for understanding the tumor microenvironment. Several studies used automated image analysis to quantify the microenvironmental composition [
20‐
23]. We previously applied automated image analysis on ovarian cancer and showed the stromal cell ratio was an independent prognostic factor [
24]. A key advantage of this method is that it is based on fully automated image analysis of H&E-stained histology slides generated as part of clinical routine, therefore does not rely on antibodies and represents an opportunity to develop cost-effective biomarkers. However, how the quantification of the microenvironmental composition can be useful beyond prognostic biomarkers and for guiding ovarian cancer treatment remained elusive. This inspired us to design our current study, in which we proposed that a simple yet effective measure of stromal cell ratio was a potential predictive biomarker for ovarian cancer platinum treatment resistance. Here, we further demonstrated the association between microenvironmental composition and platinum-resistant relapse of ovarian cancer using automated pathological image analysis as compared to established clinical variables.
One of the important findings in this study was that high stromal cell ratio as well as low cancer cell ratio based on computational analysis of routinely generated pathological sections were significantly correlated with high frequency of platinum-resistant recurrence. The majority of stromal cells identified by image analysis were fibroblasts and myofibroblasts according to α-smooth muscle actin (α-SMA) staining which has been reported in our previous study [
24]. Mhawech-Fauceglia et al. [
29] revealed that the level of fibroblast activation protein alpha (FAP) expressed by fibroblasts in tumor stroma was associated with platinum resistance and shorter recurrence in ovarian cancer, which was consistent with our observation. Thus far, the mechanism underlying the acquisition of platinum resistance in stromal cells has been rarely elucidated. Carcinoma-associated fibroblasts (CAFs) are abundant and heterogeneous stromal cells in tumor microenvironment. Recently, Su et al. found that CD10
+GPR77
+CAFs could promote tumor formation and chemoresistance [
15]. They demonstrated that mechanistically, CD10
+GPR77
+CAFs were driven by persistent NF-kB activation via p65 phosphorylation and acetylation, which was maintained by complement signaling via GPR77, a C5a receptor [
15].
However, we could not stratify patients with distinctly different relapse types by pathologist’s scores for stromal, cancer cell ratio as well as lymphocyte infiltration. Such difference could be explained by a number of factors. Firstly, pathological scoring is highly important in research and clinical settings. However, for tasks without well-established guidelines, such as stromal cell scoring which is often not part of clinical assessment, it could be prone to variability. Secondly, because of the time-consuming nature of the task, which is to estimate the amount of stromal cells in whole-tumor section slides, the pathologist can only provide visual estimation based on scoring of five high-power fields. In contrast, automated image analysis enumerates all cells in the entire slides to provide quantitative counts, thereby relieving pathologists from repetitive tasks such as cell counting. The results suggested that automated image analysis by quantifying the amount of cells can generate accurate data and reveal tumor features correlating with treatment outcome, highlighting an opportunity for performing repetitive tasks using computers and providing a new method for studying platinum-resistant relapse of ovarian cancer.
In this study, multivariate logistic regression revealed the stromal cell ratio based on automated image analysis and normalization of CA125 level after 3 cycles of chemotherapy were the significant predictors of platinum-resistant recurrent ovarian cancer. CA125 is a widely used biomarker of ovarian cancer for evaluating disease and monitoring the treatment efficacy [
30,
31]. Feng et al. found that increased preoperative CA125 level was associated with a low prognostic nutritional index (PNI), which could reflect both the nutritional and immunological status of the host [
32]. Furthermore, Miao et al. reported that the PNI could help predict the platinum resistance of ovarian cancer [
33]. However, limited researches investigated the direct association between the CA125 level and the platinum resistance of ovarian cancer. In the current study, the AUC-ROC of model including normalization of CA125 level after 3 cycles of chemotherapy was 0.79, with a sensitivity of 72.70% and specificity of 85.70%. It was noteworthy that the AUC-ROC was increased to 0.89 when the model containing stromal cell ratio combined with normalization of CA125 level. Although the increased AUC of two-parameter model was only marginally significant comparing with stromal cell ratio, and achieved a very small trend toward significant difference comparing with CA125 level, we considered that the number of subjects was too small to reach significant difference. Our findings implied that stromal cell ratio may be used to improve the predictive model when combined with a commonly used biomarker CA125. While this remains to be validated, early detection of high-risk patients could have significant implications for improved clinical management of ovarian cancer.
There were a number of limitations of our study, including that the lack of validation cohorts that have all these data types for this relative rare cancer type, the need for antibody staining to more closely examine diverse stromal cell populations and the need to understand the mechanisms underlying platinum resistance. Moreover, it should be noted that our data were restricted in patients with recurrent ovarian cancer, which might limit the generalizability of our results to broad population.
Acknowledgements
The authors gratefully acknowledge Dr. Chunbo Xie, Sun Yat-sen University Cancer Centre, Guangzhou, China, who offered help in statistical analyses.