CT radiomics prediction of CXCL9 expression and survival in ovarian cancer

Ovarian cancer (OC) has the highest mortality rate among all gynecological tumors, with no early symptoms, owing to the deep pelvic location of the ovaries and broad drug resistance. Despite advances in treatments including surgery, chemotherapy, targeted therapy, and immunotherapy, the overall survival (OS) rate of patients with OC remains low; five-year survival is less than 30% and a three-year recurrence rate is more than 70% [1, 2]. Furthermore, owing to high tumor heterogeneity, classic biomarkers and imaging indicators such as serum CA125 and transvaginal ultrasound are insufficient for monitoring therapy, which may lead to misdiagnosis or overtreatment [3, 4]. Therefore, new prognostic markers must be explored to provide individualized precision treatments.

In addition to malignant tumor cells, tumors contain normal cells, including immune cells, fibroblasts, and epithelial cells. The tumor immune microenvironment (TIME) is composed of these cells, and inflammatory immune cells act as the initial line of immune protection against pathogens [5, 6]. The mobilization of lymphocytes, a symptom of inflammation and a characteristic feature of malignancy, requires a multitude of cytokines and stimulating agents [7]. Chemokines, as a type of cytokines in the TIME, may be associated with patient outcomes. The CXC chemokine subfamily member CXCL9 encodes secreted proteins that play essential roles in disease processes, such as inflammation, immune regulation, tumor metastasis, and angiogenesis [8‐10]. In addition to its two family members, CXCL10 and CXCL11, CXCL9 has been reported to enhance antitumor lymphocyte infiltration through its receptor CXCR3 in solid tumors, such as colorectal cancer, bladder cancer, gastric cancer, and uterine corpus endometrial carcinoma [11‐14]. The same is true for ovarian cancer, where preclinical models have demonstrated a positive correlation between CXCL9 expression and T cell infiltration and overall survival [15‐17]. In view of the promising clinical applications of CXCL9, recent clinical studies have focused on its role in diseases such as COVID-19, autoimmune diseases, and cancer [18‐23]. In ovarian cancer, Au et al. indicated that high levels of CXCL9 are associated with an enhanced response to chemotherapy [24]. In addition, CXCL9 may be a reliable biomarker for predicting the immune checkpoint blockade (ICB) response in patients with OC due to CXCR3 chemokine activity being essential for effective immune checkpoint suppression [22, 23].

CXCL9 expression detected in the peripheral blood may not be representative of the tumor parenchyma. Given the remarkable tumor heterogeneity in OC and the impractical and invasive procedure of repeated biopsy, the whole tumor lesion and response to therapy are difficult to assess using conventional biopsies. Computed tomography (CT), which is widely used in clinical practice, is a common imaging method for OC diagnosis, treatment evaluation, and postoperative follow-up. Notably, rapid advances in artificial intelligence mean that radiomics, a high-throughput data mining approach that extracts massive image parameters, can now dynamically, noninvasively, and quantitatively assess the entire three-dimensional tumor [25, 26]. Previous reports have suggested that radiomics can be utilized not only in diagnosis for early OC, subtype classification, and lymph node metastasis, but also for the assessment of residual lesions, tumor heterogeneity, and TIME [27‐32]. However, CXCL9 expression has not yet been predicted using radiomics in patients with OC.

Therefore, we developed a radiomic model using the TCGA and TCIA databases to predict CXCL9 expression in patients with OC and explored its prognostic value.

Despite significant advances in cytoreductive surgery and systemic therapies, the survival rates for ovarian cancer are still vastly variable, owing to the considerable heterogeneity. Maximum tumor resection is believed to improve the prognosis of patients with ovarian cancer. Furthermore, for patients in advanced stages of the disease, aggressive treatment does not significantly extend patient survival, but reduces the quality of life [3]. Thus, additional prognostic information enables a better clinical decision prior to treatment. Increasing evidence has shown that overexpression of CXCL9 mediates the recruitment of tumor-targeted CXCR3 + T cells and natural killer (NK) cells in various solid cancers, and thereby suppresses tumor growth [8]. Studies on OC, although limited, have indicated that significant increases in CXCL9 levels are strong and independent predictors of improved survival in patients with OC [16, 22‐24]. Our results agree well with those of previous studies and demonstrate the importance of CXCL9 in the survival of patients with OC. Specifically, parameters including age, tumor histological grade, and lymphatic invasion were included in both the univariate and multivariate Cox regression analyses, and the results revealed that high CXCL9 expression levels were protective for patient survival (HR = 0.56, 95% CI: 0.417–0.75, P < 0.001).

Radiomics is a rapidly advancing quantitative technique that attempts to capture tumor characteristics using advanced imaging features. Numerous studies have shown that radiomics can reveal tumor heterogeneity and the underlying genomic and biological characteristics [26]. The same is true for ovarian cancer, favorable radiomics applications in OC like early diagnosis, subtype classification, treatment response prediction, lymph node metastasis, and survival [29‐32], raised the possibility of radiomics uncovering the CXCL9 status in OC patients.Therefore, we aimed to apply radiomics to the prediction of OC survival using CT features extracted from pretreatment images of primary tumors. So far as we know, the current study first report a radiomic model to noninvasively predicting CXCL9 expression in OC. The results revealed that the survival rate in the high rad_score group was consistently higher (P < 0.05). Considering the previous results, we believe that predicting CXCL9 expression levels based on radiomics would be helpful for clinical decision-making and individualized treatment.

Recently, deep learning radiomics integrated with RNA-Seq microenvironmental analysis has provided powerful insights into the molecular mechanisms of cancer, and contributed to survival prediction. Zhao [34] conducted a radiomics study to predict the Epstein-Barr virus status in gastric cancer, and the predictive power of the model was excellent in both the training and validation cohorts, which achieved AUCs of 0.919 and 0.939, respectively. Lu [35] used a radiomic model to predict EGFR status in lung cancer, and the prediction model using different features achieved AUCs of 0.68, 0.67, and 0.69. Moreover, an increasing number of researchers are investigating accurate and cost-effective methods to assess immune biomarkers with prognostic value in OC, among which radiomic models are becoming a trend for future studies. For example, Wan [31] built a radiomic model for C–C motif chemokine receptor type 5 (CCR5) status and OC survival analysis, which yielded an AUCs of 0.770. Gao’s radiomic model for PD-1 and OC survival analysis yielded AUCs of 0.810. Our predictive model for CXCL9 status consistently performed well in both the training and validation cohorts and achieved AUCs of 0.781 and 0.743, respectively. GLCM and GLDM textures are well-known high-order radiomic features that are highly correlated with tumorgrades [36]. In this study, we derived 5 radiomic features, among which GLCM and GLDM also played a significant role in describing the spatial relationships between pixels and distinctive heterogeneity, and reflected the expression levels of CXCL9. We further assessed the calibration ability and clinical efficacy of our model in predicting CXCL9 status in both the training and validation groups, and the results were consistent with other studies that demonstrated that radiomic models performed well in the survival prediction of OC patients [37].

Recently, some researchers have used CT texture analysis to predict treatment response and prognosis in patients with hepatic cancer [38] and glioblastoma [39]. Preoperative enhanced CT texture analysis helps predict complete response to treatment. However, correlating individual texture features with complex tumor biological processes remains a challenge because it is not possible to fully exploit meaningful clinical information for comprehensive analysis, nor to validate the reliability of the results with internal or external data. Therefore, it is common to construct multifactorial radiomics models to evaluate clinical outcomes [40]. In this study, we developed prognostic models using integrated clinical-radiogenomic information to analyze survival in patients with OC and demonstrated that RFs combined with clinical features can improve the accuracy of individual clinical decision-making, and thus has a high potential for evidence-based clinical translation for OC management.

However, given the exploratory nature of this retrospective study, future studies are required to validate the findings in this study. The limited number of cases is one of the main factors restricting our research from reaching reliable conclusions. Although tenfold internal cross-validation was used for the evaluation of signature model, lack of external cohort from OC patients is another limitation of the present study. In addition, radiomics studies have generally demonstrated predictive ability in cancers, but the sensitivity and specificity of the predictions were poorer using different models and radiological signatures. Additionally, free datasets can vary in CT image quality and show high imbalances, especially in the ratio of high versus low CXCL9 levels. Accordingly, larger cohort data from multiple centers may provide a solution for future research and result in practical applications.

In conclusion, our model can be used to evaluate the prognostic risk in patients with OC. The radiomics-based nomogram, which incorporates clinical and CT characteristics, can predict CXCL9 status in a noninvasive manner, potentially fulfilling the ultimate purpose of precision medicine for ovarian cancer.