Introduction
Ovarian cancer is the most lethal gynecological cancer worldwide [1], with high‐grade serous ovarian cancer (HGSOC) making up 70%–80% of ovarian cancer‐related deaths [2]. Currently, aggressive cytoreductive surgery and platinum‐based chemotherapy are the standard first‐line treatments for this disease [3, 4]. For about 25% of patients, first‐line chemotherapy does not stop disease progression, substantially reducing survival [3].
Current methods for identifying patients at risk for platinum resistance are inadequate, highlighting the need for more effective predictive biomarkers [5, 6]. Standard pretreatment CT imaging may serve as a direct prognostic tool for the development of platinum resistance.
Progression‐free survival (PFS), the time from diagnosis to disease progression or death, is critical in assessing sensitivity to platinum‐based chemotherapy. Patients who relapse more than 6 months after completing primary platinum‐based treatment are considered platinum‐sensitive, with most responding to subsequent lines of treatment [3]. In contrast, those with a PFS of less than 6 months are classified as platinum‐resistant and exhibit significantly lower response rates to future platinum‐based treatments [3].
A subset of patients with platinum‐resistant disease, described as platinum‐refractory, experience progression during treatment and face the poorest prognosis, with a median overall survival of approximately 3–5 months [3].
Most research on platinum resistance and PFS has explored genomic biomarkers [7, 8], tumor proteomic signatures [9], and tumor microenvironments [10], aiming to capture the biological mechanism underlying treatment response. Recent research in ovarian cancer has focused on CT‐based body composition as a predictor for outcomes, including survival [11, 12, 13, 14, 15, 16, 17, 18, 19, 20] and chemotoxicity [11, 19, 21, 22, 23, 24, 25, 26]. We identified only three studies [14, 19, 27] that evaluated the relationship between CT‐based body composition and platinum resistance specifically.
Kim et al. [14] examined 179 patients with HGSOC and assessed muscle mass metrics based on the cross‐sectional area of skeletal muscle at the L3 level, finding no significant relationship with platinum resistance. Similarly, Bruno et al. [19] analyzed L3‐based muscle and adipose tissue measurements in 239 patients and reported no association with platinum resistance. Aichi et al. [27] conducted a retrospective analysis of 52 patients with platinum‐resistant recurrent ovarian cancer and found that low L3‐based muscle mass measurements were significantly associated with shorter survival (HR 1.93, p = 0.03).
A significant limitation of these studies is relying on L3‐based CT metrics computed from a single 2D‐abdominal axial image slice at L3 to estimate whole‐body composition. Although L3‐based CT analysis is considered the reference standard for body composition assessment, single‐slice mid‐L3 measurements remain susceptible to error due to sex‐related differences in body composition, variability in L3 slice selection across studies, and inconsistencies among automated segmentation algorithms [28, 29].
Our study aimed to investigate the association of pretreatment clinical markers and CT abdominal body composition features with platinum resistance and PFS. The abdominal region was examined due to the availability of routine CT imaging with an abdominal view and to capture visceral, subcutaneous, and intramuscular adiposity. We used automatic body composition segmentation with deep learning rather than relying solely on L3‐based measurements to enable more accurate analyses.
To capture a comprehensive body composition profile, we evaluated five compartments encompassing skeletal muscle, bone, and three adipose tissue depots (visceral, subcutaneous, and intramuscular fat). In addition to volumetric measurements, we assessed tissue radiodensity (Hounsfield units), which may provide complementary information on tissue quality, including fatty infiltration of skeletal muscle, bone mineralization, and other compositional changes not reflected by volume alone. The ratio between adipose tissue and muscle volumes was added to the analysis based on prior studies correlating high fat‐to‐muscle ratio with lower survival in patients with gynecological cancer [12, 30].
Using various survival models, we also evaluated the association of these CT body composition features with PFS and built models using different combinations of features. Finally, we hypothesized that composite CT‐Clinical models, trained on both CT and clinical variables, would provide better prognostic modeling than clinical‐only models.
Materials and Methods
Study Cohorts
Internal Cohort
We obtained data on 532 women treated for primary ovarian cancer at the Hillman Comprehensive Cancer Center from 2011 to 2023 and followed through September 2024. Women who did not have HGSOC (N = 204) and did not receive platinum‐based chemotherapy (N = 7) were excluded, leaving 321 women with HGSOC for the clinical analysis. For the body composition analysis, cases with no pretreatment scans, scans with fewer than 50 slices, or scans with slice thickness greater than 5 mm (N = 91) were excluded, leaving 230 patients with usable pretreatment CT imaging with an abdominal field‐of‐view (Figure 1).
Study participants were originally consented into a data bank and biorepository (UPCI 07‐058), where informed consent was obtained to release de‐identified data and specimens to investigators with IRB‐approved protocols. As all data were de‐identified via an honest broker prior to sharing with the investigative team, this study received exempt IRB approval from the University of Pittsburgh Institutional Review Board (IRB 20120024). The data that support the findings of this study are available from the corresponding author upon reasonable request.
External Validation Cohort
Clinical data and CT images for the TCGA‐OV cohort [31] of 488 ovarian serous cystadenocarcinoma cases were obtained through the cBioPortal website [32] and The Cancer Imaging Archive portal (https://www.cancerimagingarchive.net/) [33], respectively. The demographics and clinical data of this cohort are described in Table S1. We excluded 177 cases missing platinum‐resistance data, 68 non‐HGSOC cases, and 185 cases with missing or low‐quality CT scans (< 50 slices, > 5 mm thickness), leaving 58 cases for analysis (Figure 1).
CT Image Features
Five body composition features—visceral adipose tissue (VAT), subcutaneous adipose tissue (SAT), intramuscular adipose tissue (IMAT), skeletal muscle, and bone—were automatically segmented for each CT scan using our validated deep learning algorithms [34, 35]. To normalize anatomical variations across individuals and address CT scan heterogeneity, body composition tissues spanning the top and bottom of the individual's abdomen were quantified (Figure 2). Within our cohort, CT scans for clinical assessment used either 64 detector row CT scanners LightSpeed VCT (GE Medical Systems, Milwaukee, WI) or LightSpeed Pro 16 (GE Medical Systems, Milwaukee, WI) with the following parameters: automatic tube current ranging from 150 mA to 600 mA, matrix of 512 × 512, section thickness of 0.625 mm to 5.0 mm, 64 × 0.625‐mm detector collimation, and spiral pitch factor of 1.375.
Contrast‐enhanced CT images were obtained after intravenous injection of nonionic iodinated contrast medium ISOVUE 300.
Body composition tissues were segmented and quantified using an in‐house 3‐dimensional convolutional neural network algorithm with a grid‐based three‐dimensional patch sampling operation [35]. Based on these segmentations, each tissue's volume was computed. We also computed their average densities expressed in Hounsfield Units (HU) for body tissues.
While volumetric measures quantify the amount of tissue present, HU density reflects tissue composition and quality; for example, lower muscle density may indicate greater fatty infiltration, whereas lower bone density may reflect reduced bone mineralization. Importantly, for adipose tissue, higher density corresponds to less negative HU values, whereas for bone and muscle, higher HU values generally indicate denser tissue composition. Ratios between tissue types, such as the fat‐to‐muscle ratio (FMR), were calculated by dividing the segmented volumes.
These algorithms have been validated across multiple studies and perform reliably on various CT protocols and scanners [34, 35].
Clinical Features
Clinical characteristics were extracted from medical records by staff trained in medical records abstraction and verified by comparison to the Hillman Cancer Center cancer registry, which is accredited by the American College of Surgeons Commission on Cancer. Differences between the two data sources were discussed by the study team with a review of the electronic medical records, and the final classification was determined by the study gynecologic oncologist or pathologist, as appropriate. A single pathologist reviewed all patient primary diagnosis slides to confirm histology.
Differences between the original pathologist's classification and pathology review were resolved by the original pathologist and study pathologist together, or if the original pathologist was not available, with an independent pathologist who provided consultation as a courtesy to our study.
Extracted clinical data included basic demographics, smoking history, date of diagnosis, date of the last chemotherapy infusion, date of first progression or recurrence, pre‐ and post‐treatment CA‐125 levels, pretreatment neutrophil, lymphocyte, and platelet counts, FIGO stage (I–IV), and optimal debulking status (residual disease < 1 cm). Patients were also classified according to chemotherapy sequence (neoadjuvant or adjuvant only) and platinum sensitivity under RECIST 1.1/GCIG criteria. Platinum‐resistant disease was defined as progression during treatment or within 6 months after the last chemotherapy infusion.
In addition to standard clinical covariates (e.g., age, stage, histology), pretreatment lab values were extracted, including neutrophil, lymphocyte, and platelet counts, which were used to derive two established prognostic markers of survival and disease progression [36, 37]: the neutrophil‐to‐lymphocyte ratio (NLR) and systemic immune‐inflammation index (SII). The SII was calculated by multiplying the NLR by the platelet count [37].
PFS was defined as the time (in months) between diagnosis and date of first progression (for patients who progressed) or date last known to be alive (for patients who did not progress).
Statistical Analysis
Univariate and multivariate platinum resistance analyses were conducted separately for NACT and adjuvant treatment groups to assess treatment‐specific associations. For PFS analyses, patients from both treatment groups were pooled due to limited event counts within individual subgroups.
Platinum Sensitivity Analysis
Associations between platinum resistance and both clinical and CT‐derived body composition variables were evaluated. Body composition variables included VAT, SAT, IMAT, skeletal muscle, and bone, with each tissue assessed using both volumetric and radiodensity (HU) measurements. All body composition features were analyzed as continuous variables without predefined cutoffs for “high” or “low” values.
Categorical variables were compared between platinum‐sensitive and platinum‐resistant groups using the chi‐square test or Fisher's exact test, as appropriate, while continuous variables were compared using the Mann–Whitney U test. Separate analyses were conducted for adjuvant and NACT treatment groups. Variables demonstrating evidence of association in univariate analysis (p < 0.10) were subsequently entered into multivariable logistic regression models to predict platinum resistance.
Multicollinearity was addressed by removing features with a variance inflation factor (VIF) greater than five. NACT versus adjuvant treatment was adjusted for in multivariable models. Continuous variables exhibiting substantial skew were log‐transformed prior to analysis, and categorical variables were encoded using one‐hot encoding.
After univariate analysis, the clinical multivariate logistic regression model was constructed using six significant clinical features: age, pretreatment CA125, SII, NLR, optimal debulking status, and advanced stage. Age and NLR were excluded from the final analysis due to high multicollinearity. To improve the linearity and stability of the model, pretreatment CA125 and SII values were log‐transformed.
In the CT‐based multivariate logistic regression model, muscle density (HU), IMAT volume (mL), and SAT density (HU) were chosen based on their univariate associations with resistant disease.
PFS Analysis
Clinical variables and CT‐derived body composition features assessed in the platinum resistance analysis (including volumetric and radiodensity measures of VAT, SAT, IMAT, skeletal muscle, and bone) were evaluated for their association with PFS using Cox proportional hazards regression. All body composition variables were modeled as continuous predictors without predefined cutoffs. Univariate hazard ratios were estimated for each feature individually, and variables demonstrating evidence of association (p < 0.10) were considered for multivariable modeling.
The proportional hazards assumption was assessed for all covariates prior to model fitting. Continuous variables exhibiting substantial skew were log‐transformed to improve linearity and model stability. To reduce multicollinearity, highly correlated variables were excluded from multivariable models based on domain knowledge and VIF > 5.
Multivariable models were constructed separately for clinical and CT‐derived feature groups and adjusted for chemotherapy type (NACT vs. adjuvant). A two‐sided p‐value < 0.05 was considered statistically significant.
PFS Modeling
Pretreatment variables demonstrating a univariate association with PFS on Cox analysis (p < 0.10) were retained for multivariable modeling. Continuous variables were standardized using z‐score normalization, and CA125 and SII were log‐transformed before standardization to account for right‐skewed distributions. Missing values were imputed using median imputation for continuous variables and mode imputation for categorical variables.
The prediction window was set to 5 years, capturing the majority of available follow‐up data (median PFS: 613 days, IQR: 358–1219).
Three prespecified feature sets were evaluated: (1) chemotherapy sequence (NACT vs. adjuvant) as a baseline reference, (2) clinical features only (age, CA125, SII, optimal debulking status, disease stage, chemotherapy type), and (3) clinical features combined with CT‐derived abdominal body composition features. Each feature set was evaluated against five model classes: Elastic Net Cox (l1_ratio = 0.1, predominantly ridge‐penalized), ridge Cox (l1_ratio = 0.0, pure L2 penalty), survival support vector machine (SVM) with RBF kernel, PCA‐reduced Cox proportional hazards (retaining 95% of variance), and standard unpenalized Cox proportional hazards. The Elastic Net and ridge Cox models were fitted using automatic alpha selection via internal cross‐validation.
A stacked ensemble was additionally evaluated, in which out‐of‐fold risk scores were generated separately from the clinical‐only and abdominal‐only elastic net Cox models using 10‐fold cross‐validation, and a meta elastic net Cox model was trained on the two resulting risk scores. This approach was used to assess whether clinical and CT‐derived prognostic signals were independently complementary.
Model discrimination was assessed using the time‐dependent cumulative AUC, estimated within the interval supported by both training and test data in each fold to prevent extrapolation beyond observed follow‐up. Model performance was evaluated using 10‐fold cross‐validation. To assess whether the addition of CT‐derived abdominal body composition features significantly improved model discrimination over clinical features alone, paired statistical comparisons were performed between the clinical and combined CT and clinical feature sets for each model.
Identical 10‐fold cross‐validation splits were used to generate fold‐level time‐dependent AUC scores for both feature sets, ensuring each fold‐level difference represented a controlled paired observation on the same held‐out patients. Given the small number of paired observations (10 folds) and the absence of an assumption of normally distributed differences, the Wilcoxon signed‐rank test was applied. To account for multiple comparisons across model classes, p‐values were corrected using the Benjamini‐Hochberg false discovery rate (FDR) procedure.
An FDR‐corrected p‐value below 0.05 was considered indicative of a statistically significant improvement in discrimination attributable to the CT body composition features. All analyses were conducted in Python 3.9.1.
Results
Clinical Features and Platinum Sensitivity
Among the 321 study subjects with HGSOC who underwent platinum‐based chemotherapy, 90 (28.0%) had platinum‐resistant disease (Table 1). As expected, patients with platinum‐resistant disease had significantly worse overall and progression‐free survival (Figure S1). Older patients were more likely to be platinum‐resistant (age 66.9 ± 8.7 vs.
64.6 ± 10.1 years; p = 0.039). Established clinical variables associated with platinum resistance included advanced stage (97.6% vs. 74.7%; p < 0.001), nonoptimal debulking (16.7% vs.
7.0%; p = 0.018), and the use of NACT (73.3% vs. 35.9%; p < 0.001). Additionally, increased pretreatment CA125 levels were significantly associated with resistance (p < 0.001), as were several inflammatory markers, including increased platelet count (p = 0.009), NLR (p = 0.022), and SII (p = 0.005) (Table 1).
CT Features and Platinum Sensitivity
In univariate analysis, several abdominal CT‐derived body composition features were associated with platinum resistance (Table 2). In the adjuvant‐only cohort, platinum‐resistant patients demonstrated lower muscle density (32.8 ± 11.8 vs. 39.5 ± 11.3 HU, p = 0.025), higher IMAT volume (0.569 ± 0.178 vs.
0.487 ± 0.231 mL, p = 0.040), and higher SAT density (−82.9 ± 23.6 vs. −91.6 ± 16.6 HU, p = 0.047) compared with platinum‐sensitive patients. In the neoadjuvant chemotherapy cohort, platinum‐resistant patients had lower bone density (306.3 ± 64.4 vs. 324.1 ± 62.2 HU, p = 0.034), while other muscle and adipose parameters were not significantly associated with platinum resistance.
In the combined cohort, platinum resistance remained associated with lower muscle density (34.8 ± 12.9 vs. 38.6 ± 11.3 HU, p = 0.012), higher IMAT volume (0.602 ± 0.283 vs. 0.511 ± 0.238 mL, p = 0.017), and higher SAT density (−86.5 ± 20.4 vs. −91.0 ± 16.7 HU, p = 0.031).
In the external validation cohort, higher SAT density (−85.0 ± 15.5 vs. −92.1 ± 13.0 HU, p = 0.034) was similarly associated with platinum resistance (Table S2).
Platinum Sensitivity Multivariate Analysis
In the multivariate clinical model, log‐pretreatment CA125 (OR = 1.36, p = 0.001) and advanced stage (OR = 10.03, p = 0.002) remained significant independent predictors of resistance. Adjusting for NACT, advanced stage (OR = 7.41, p = 0.008), optimal debulking status (OR = 0.39, p = 0.027), and log‐pretreatment CA125 (OR = 1.24, p = 0.029) maintained statistical significance as independent predictors of platinum‐resistant disease (Table 1).
In the CT‐based multivariate model using abdominal features, only SAT density remained a significant independent predictor of resistance (OR = 1.02, p = 0.016), while abdominal muscle density and IMAT volume did not reach statistical significance. After adjusting for NACT sequence, abdominal SAT density maintained its independent statistical significance (OR = 1.02, p = 0.022), while abdominal muscle density and IMAT volume remained non‐significant predictors (Table 2).
Clinical Features and PFS
In univariate Cox proportional hazards analysis, several clinical variables were significantly associated with PFS (Table 3). Higher preoperative CA‐125 levels were associated with an increased risk of earlier progression (HR 1.18, 95% CI: 1.09–1.29, p < 0.001), as was advanced disease stage (HR 3.19, 95% CI: 2.12–4.79, p < 0.001). Increasing age was also associated with shorter PFS (HR 1.01 per year, 95% CI: 1.00–1.03, p = 0.009).
In contrast, optimal debulking was protective and associated with prolonged PFS (HR 0.61, 95% CI: 0.40–0.94, p = 0.025). In group multivariate analysis with all clinical features, age, CA‐125, and advanced stage remained significant, while advanced stage only remained significant when adjusting for NACT sequence.
CT Features and PFS
Multiple abdominal body composition features were associated with PFS. Increased IMAT volume was strongly associated with a higher risk of earlier progression (HR 2.83, 95% CI: 1.62–4.95, p < 0.001), while lower skeletal muscle density was associated with worse PFS (HR 0.98 per HU, 95% CI 0.97–0.99, p = 0.002). Higher visceral, subcutaneous, and cavity volumes were each associated with increased risk of progression, with hazard ratios ranging from 1.05 to 1.20 (all p < 0.05).
No features were significant in multivariate group analysis or when adjusting for NACT (Table 3).
PFS Models
Model discrimination results are presented in Table 4. The NACT‐only baseline model achieved a mean AUC of 0.644 (95% CI: 0.567–0.721). Among clinical‐only models, PCA‐Cox performed best (AUC: 0.677, 95% CI: 0.612–0.743), followed by Elastic Net Cox (AUC: 0.673, 95% CI: 0.607–0.738) and standard CoxPH (AUC: 0.669, 95% CI: 0.601–0.737).
Ridge Cox and SVM performed notably worse (AUC: 0.573 and 0.570, respectively).
The addition of CT‐derived abdominal body composition features consistently improved discrimination across model classes. The Elastic Net Cox model achieved the highest overall AUC of 0.704 (95% CI: 0.660–0.748), representing an improvement over the clinical‐only Elastic Net model. The Ridge Cox (AUC: 0.695, 95% CI: 0.640–0.750) and SVM (AUC: 0.629, 95% CI: 0.587–0.671) models both showed significant improvement with the addition of CT features (p = 0.011 and 0.043, respectively).
The unpenalized CoxPH model could not be evaluated for the combined feature set due to violations of modeling assumptions. The stacked ensemble, combining out‐of‐fold clinical and abdominal risk scores as meta‐features, achieved an AUC of 0.677 (95% CI: 0.614–0.741), which did not exceed the best single combined model, suggesting limited independent complementarity between the two risk scores at this sample size.
Overall, the Elastic Net Cox model trained on combined clinical and CT abdominal features demonstrated the best discrimination, supporting the incremental prognostic value of pretreatment body composition over clinical variables alone (Table 4).
Discussion
Summary of Main Results
This study examined the relationship between clinical factors, CT‐based body composition features, and both platinum resistance and PFS in women with HGSOC. Current methods to predict these critical clinical outcomes are limited and largely do not consider body composition [3, 4]. Using previously developed deep learning algorithms, we automatically segmented key body tissues, including VAT, SAT, IMAT, skeletal muscle, and bone.
We identified multiple CT features associated with both platinum resistance and PFS and partially validated our findings on an external cohort. Overall, low muscle density and high adiposity, specifically subcutaneous and intramuscular, were consistently associated with both platinum resistance and shorter PFS.
Multiple survival models were then developed to integrate clinical factors and CT‐derived body composition features to predict PFS using 10‐fold cross‐validation. Two models significantly improved in their PFS prediction with the addition of body composition features, and the best model (Elastic Net Cox) achieved an AUC of 0.704 (95% CI, 0.660–0.748).
Results in the Context of Published Literature
CT imaging is extensively used in clinical practice to manage ovarian cancer [3]. It provides exceptional resolution for visualizing various body tissues, enabling the exploration of biomarkers potentially linked to chemotherapy response [19, 22, 30]. Body composition, which includes muscle, bone, and various adipose tissues, can offer deeper insights into nutritional status beyond traditional BMI measurements [35].
While prior ovarian cancer studies have largely evaluated body composition as a prognostic marker of survival or postoperative outcomes, relatively few have examined its utility for predicting platinum resistance, and even fewer have externally validated these associations [22, 30]. Our study extends the existing literature by evaluating comprehensive 3D measurements across multiple adipose and muscle compartments and demonstrating reproducible associations for several body composition features across independent cohorts.
Among the previous studies on ovarian cancer and body composition, the vast majority use L3 single‐slice segmentation to estimate full‐body tissue composition [11, 12, 13, 16, 18, 19, 20, 22, 25, 38]. Often, multiple image slices encompass the L3 vertebrae, so choosing a single slice may inadequately estimate total body composition, limiting study findings [28, 29]. Additionally, most studies focus on skeletal muscle index and other muscle‐related metrics [11, 12, 13, 14, 16, 18, 19, 20, 22, 23, 25].
However, the threshold values used to define these indices are often arbitrary, derived from the specific data distribution of each study, and lack standardized consensus across the literature [13, 22, 28]. Accordingly, we chose not to apply muscle index‐classifications and instead report absolute 3D muscle volumes and density measurements to provide a more objective assessment.
Strengths and Weaknesses
We analyzed the issue of platinum resistance and PFS from a physiological angle, evaluating five different types of body composition. We found that certain CT body features were consistently associated with platinum resistance and PFS. Increased SAT density was correlated with resistance in both internal and external cohorts.
This follows previous studies suggesting that SAT enhances chemotherapeutic drug resistance by activating specific signaling pathways [39]. Similarly, increased IMAT was associated with both platinum resistance and shorter PFS. High IMAT reflects muscle wasting, sarcopenia, and chronic inflammation and has been associated with oncological outcomes in prior studies [40].
In our study, VAT measurements were associated with PFS, a finding that follows established research on how visceral adiposity contributes to ovarian cancer progression through hormonal and metabolic pathways [41].
Adipose and muscle compartments are increasingly recognized as markers of physiologic reserve, systemic inflammation, and metabolic dysfunction, all of which can influence a patient's ability to tolerate chemotherapy [40, 42]. Increased adiposity has been linked to chronic inflammatory signaling, altered drug pharmacokinetics, insulin resistance, and tumor‐promoting endocrine pathways [41]. These factors may contribute not only to reduced treatment efficacy but also to poorer chemotherapy tolerance, increased treatment‐related toxicity, dose reductions or delays, and ultimately inferior survival outcomes [21, 40, 41, 42].
Consequently, body composition biomarkers may provide clinically relevant information by identifying patients at increased risk of platinum resistance before treatment initiation [37]. If validated prospectively, body composition biomarkers could complement established clinical and molecular predictors to improve individualized treatment planning, guide supportive interventions such as nutritional optimization and exercise programs, and identify patients who may benefit from closer monitoring or consideration of alternative therapeutic strategies [3, 4, 22, 30, 38, 43].
One limitation of the study was that the internal cohort mainly consisted of women who self‐identified as White. This precludes generalizing our results to women of other races. Ongoing work will validate these results on more diverse populations.
Another limitation was that the external cohort had limited clinical features for analysis. As a result, external validation was largely limited to imaging‐derived body composition features, and many of the clinical characteristics associated with platinum resistance in the internal cohort could not be systematically evaluated or compared. Consequently, we were unable to fully assess the reproducibility of the clinical and combined clinical‐radiographic associations observed in our institutional dataset.
Additionally, adjuvant and neoadjuvant cases could not be completely assessed independently using survival models due to low progression events in the adjuvant cohort. We acknowledge that patients who undergo NACT have advanced disease processes and may benefit from separate analyses. Finally, while our models included some biological prognostic markers derived from medical record data (e.g., NLR, SII), we could not include proteomic, genomic, and other tumor microenvironment markers that may also impact platinum resistance and progression.
Implications for Practice and Future Research
Exploring our findings in contemporaneous cohorts with more racially diverse populations and incorporating additional clinically evaluated biological markers, such as tumor immune markers, may help improve predictions and provide a valuable tool for clinicians to identify women with a high likelihood of failing conventional treatment.
Conclusion
Platinum resistance and disease progression are strongly associated with poor ovarian cancer outcomes. Identifying women at increased risk of platinum resistance before treatment may improve patient counseling, guide surveillance strategies, and help prioritize enrollment into clinical trials or alternative therapeutic approaches. In this study, we examined platinum resistance and PFS from a joint clinical‐radiographic angle, identifying CT body composition features alongside traditional clinical variables that may serve as prognostic factors for both outcomes.
Importantly, unlike many established clinical risk factors, body composition is potentially modifiable. As such, these features may not only improve risk prediction but also help identify patients who could benefit from targeted interventions aimed at improving muscle quality, reducing adverse adiposity, or addressing frailty prior to or during treatment. Future studies are needed to determine whether modifying these body composition characteristics can translate into improved treatment response and survival outcomes.
Author Contributions
Jatin Singh: writing – original draft, writing – review and editing, conceptualization, investigation, visualization. Xin Meng: writing – review and editing, investigation. Grant Kokenberger: writing – review and editing, investigation.
Jianbing Zhu: writing – review and editing, investigation. Lauren Borho: writing – review and editing, data curation, investigation, project administration. Esther Elishaev: writing – review and editing, data curation.
Sarah Taylor: writing – review and editing, data curation. Francesmary Modugno: writing – review and editing, writing – original draft, supervision, funding acquisition, conceptualization, data curation, resources, visualization. Jiantao Pu: writing – review and editing, writing – original draft, supervision, funding acquisition, conceptualization, resources, visualization, validation.
Funding
This work was supported by the Congressionally Directed Medical Research Program (W81XWH2110338); the National Institutes of Health (R01CA237277, R01HL174570, R01HL187015, R21CA267050); and the Janet Burroughs Ovarian Cancer Foundation. Work performed in the Pitt Biospecimen Core (RRID:SCR_025229) and services and instruments used in this project were supported, in part, by the University of Pittsburgh, the Office of the Senior Vice Chancellor for Health Sciences. The results published here are partly based upon data generated by the TCGA Research Network: https://www.cancer.gov/tcga.
The funders had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.
Ethics Statement
Study participants were originally consented into a data bank and biorepository (UPCI 07‐058), where informed consent was obtained to release de‐identified data and specimens to investigators with IRB‐approved protocols. As all data were de‐identified via an honest broker prior to sharing with the investigative team, this study received exempt IRB approval from the University of Pittsburgh Institutional Review Board (IRB 20120024).
Conflicts of Interest
The authors declare no conflicts of interest.
Supporting information