Development and validation of a ¹⁸F-FDG PET/CT radiomics nomogram for predicting progression free survival in locally advanced cervical cancer: a retrospective multicenter study

In this study, we successfully developed a combined risk stratification model that incorporated clinical, PET radiomic, and CT radiomic features for individual prediction of 3- and 5-year PFS probability in patients with LACC who received definitive CCRT. The findings were further validated in an external validation cohort, and an easy-to-use nomogram was developed to aid clinical decision-making.

Several studies have demonstrated that combining radiomics and clinical features may enhance predictive performance for prognosis and therapeutic effects in cervical cancer [23, 24]; these findings are supported by the results of our study. A previous study using radiomics found older age, as defined by an age of 55 years or more, to be an adverse prognostic factor [25]. Our data demonstrated no significant correlation with age; this is in agreement with the findings from another study [26]. Clinical studies have found the pre-treatment neutrophil-to-lymphocyte, platelet-to-lymphocyte, and lymphocyte-to-monocyte ratios to be prognostic indicators [27‐29]. However, these studies did not demonstrate any statistically significant association between these ratios and PFS on univariate Cox regression. It is therefore necessary to define more optimal thresholds using an appreciable sample size. Clinical prognosis is most commonly predicted using FIGO staging. In this context, Wei et al. [30] used the FIGO stage and radiomic features to assess survival in patients with LACC. Mu et al. [31] and Jiang et al. [32] incorporated pelvic lymph nodes metastasis status and other clinical factors into radiomics models to improve the predictive value. However, we did not include two recognized prognostic factors (namely, FIGO stage and pelvic lymph nodes metastasis) for model construction, as these variables did not demonstrate significant correlation with PFS on multivariate Cox regression; this could be attributed to the relatively small number of patients. In the present study, pathological type was the only clinically significant feature predictive of PFS. Addition of this feature to the radiomics model enhanced predictive power; this is consistent with findings from previous research [33].

Notably, ¹⁸F-FDG PET metabolic parameters including SUV_max, SUV_mean, MTV, and TLG have unclear prognostic value in cervical cancer. In this context, the pre-therapeutic SUV_max of the primary tumor has been reported to correlate significantly with OS and event-free survival in patients with LACC who receive CCRT; it may therefore serve as a key prognostic predictor [34]. In their study, Calles-Sastre et al. [35] found the pre-treatment TLG and MTV to be independent prognostic factors for OS and recurrence-free survival in patients with advanced cervical cancer patients who underwent definitive CCRT; these variables were better than the widely-used parameter, SUV_max. The performance of the radiomics model in predicting PFS could be improved by incorporating MTV or TLG, as selected using univariate and multivariate Cox analysis [36, 37]. However, none of the parameters, including SUV_max, SUV_mean, MTV, and TLG have been used for clinical decision-making in primary cervical tumors [38]. Although univariate analysis indicated TLG and MTV to have predictive value for PFS in this study, this was not found on multivariate analysis. In addition, neither the SUV_max nor SUV_mean showed prognostic value; this may be attributed to central necrosis of tumor tissue and interference by the inflammatory response. Our results were similar to those observed by Chen et al. [33], in that the ¹⁸F-FDG PET metabolic parameters did not demonstrate independent prognostic value.

In a study where PET/CT-based radiomics was used to construct prediction models, 5 CT radiomic, 1 PET radiomic, and 6 clinical features were filtered. The radiomics model achieved better predictive performance in the training and internal validation datasets than in the clinical model [38]; this is in agreement with our results. We successfully extracted 4 radiomic features to predict PFS; these included 3 features derived from CT and 1 feature extracted from PET. The number of selected radiomic features derived from PET images was less than that obtained from CT images. This may indicate that CT images contain more prognostic information than PET images, or that PET radiomic features are more likely to be affected by scanning protocols and reconstruction parameters than CT radiomic features [39]. Notably, the maximum 2D diameter (row), as a shape feature, was included in the CT-derived features; this feature characterized the longest distance between tumor surface mesh vertices in the sagittal plane. Nevertheless, the tumor diameter showed no statistically significant association on univariate Cox regression in our study. This shows that radiomics is more precise than manual characterization. In addition, the 4 extracted radiomics features included one wavelet, one logarithm, and two original features. The wavelet and logarithm features could reflect tumor spatial heterogeneity in multiple scales; this suggests that more prognostic information may be mined via filter transformation of the original images. This further reflects the advantages of using radiomics; it offers valuable mined high-dimensional data that are difficult to sense manually. The gray-level non-uniformity (GLNU) feature represents the variability of gray-level intensity values in the image; higher values indicate greater heterogeneity in intensity values. This is in agreement with the findings from a previous study that suggested GLNU to be a poor prognostic marker for cervical cancer [40]. In our previous study using CT radiomics, we extracted maximum 2D diameter (row) and gray level size zone matrix derived from GLNU values to predict PFS in patients with esophageal squamous cell carcinoma [41]. In their study, Lucia et al. [42] found GLNU derived from gray-level run length matrix to be the only PET feature predictive of disease-free survival. The studies differed in that the GLNU was derived from gray-level run length matrix in their study, whereas we derived it from gray level size zone matrix. The radiomic features or models selected and developed in other studies were not fully transposable to our study cohort. Among the 4 features, 3 were textural; this may have helped to exhibit intratumoral heterogeneity and provided additional independent prognostic information [43].

Our models were evaluated in an independent dataset, in which the ROC curve displayed good performance, calibration curves showed good agreement, and DCA confirmed clinical utility; this is one of the strengths of our study. Our study is of clinical value, as is provides a visually quantitative nomogram to aid clinicians in their routine practice. The nomogram combined PET radiomic, CT radiomic, and pathological features, and achieved higher AUC values and better calibration than the radiomics or clinical models alone. In addition, the present study followed the TRIPOD guidelines [44]; this has been outlined in the Table S3.

Despite the favorable results observed using PET/CT-based radiomics, this study has certain limitations. First, it had a retrospective design and offers preliminary findings; in addition, some patients were followed-up for a relatively short duration. Longer follow-up is needed to further evaluate the long-term prognostic value of the established model and nomogram. Second, correlation between human papilloma virus infection and PFS could not be evaluated owing to the absence of complete records pertaining to HPV status in some patients. In addition, some patients lacked data pertaining to the degree of cellular differentiation and expression of serum tumor markers.

In this study, the predictive performance of the clinical model in the external validation cohort exceeded that in the training cohort, which affected the performance of the combined model across different datasets. Our study was conducted at two centers, located in Shandong Province and the Xinjiang Uygur Autonomous Region of China. Shandong is situated along the eastern coastal region, while Xinjiang is located on the northwestern border, close to West Asia. The latest data on cervical cancer in China shows that the incidence rate and mortality rate in the eastern region are higher than those in the western region [45]. Furthermore, the positive HPV infection rates differed between the Uygur and Han in Xinjiang, China, and the genotype distribution of infection was different [46]. The aforementioned limitations may be the key factor in solving this result. Future prospective studies including more clinical parameters and several clinical endpoints need to be performed to further validate and enhance the predictive performance.