Pancreatic ductal adenocarcinoma (PDAC) is associated with poor prognosis even without distant metastases, necessitating in-depth characterization of primary tumours for survival prediction. We assessed the feasibility of using FDG-PET/CT radiomics to predict overall survival (OS) in PDAC. This retrospective study included PDAC patients without distant metastasis who underwent FDG-PET/CT for initial staging. Primary tumours were segmented from FDG-PET/CT images, extracting 222 radiomics features. A radiomics-based risk score (Rad-score) was developed using Cox proportional hazards regression with LASSO to predict OS. The prognostic performance of the Rad-score was compared with a clinical model (demographics, disease stage, laboratory results) using Harrell's concordance index (C-index) and bootstrapping. 140 patients were included, with a mortality rate was 72.9% during follow-up (total population, 19.5 ± 19.2 months; survivors, 34.4 ± 28.8 months). Eleven radiomics features were significant for survival prediction. The Rad-score predicted OS with a C-index of 0.681 [95% CI, 0.632–0.731]. A model integrating clinical parameters and Rad-score outperformed the clinical-only model in predicting OS (C-index 0.736 [0.727–0.812] vs. 0.667 [0.648–0.750]; C-index difference 0.069 [0.028–0.117]; P < 0.001). These findings suggest that incorporating FDG-PET/CT radiomics into preexisting prognotic stratification paradiagm may enhance survival prediction in PDAC, warranting large-scale studies to confirm its applicability in clinical practice.
Article
The Value of PET/CT Radiomics for Predicting Survival Outcomes in Patients with Pancreatic Ductal Adenocarcinoma
https://doi.org/10.21203/rs.3.rs-4825555/v1
This work is licensed under a CC BY 4.0 License
You are reading this latest preprint version
FDG-PET/CT
Pancreas cancer
Radiomics
Survival
Pancreatic ductal adenocarcinoma (PDAC) is often associated with a poor prognosis due to its aggressive nature and resistance to conventional therapies1. Nearly half of newly diagnosed patients with PDAC present with distant metastases at the time of primary tumour detection due to early dissemination and high risk of metastasis2. PDAC tumours often exhibit significant heterogeneity, rendering them refractory to standard chemotherapy and targeted interventions. Despite advancements in surgical techniques, radiation, and chemotherapy, the 5-year survival rate for PDAC remains remarkably low, typically ranging between 10% and 15%3. Survival outcomes vary contingent on factors such as the stage of cancer at diagnosis and individual patient characteristics. Consequently, it is important to identify specific tumour characteristics associated with poor prognosis.
Positron Emission Tomography (PET) using F-18 fluorodeoxyglucose (FDG) is a valuable tool in oncology, providing functional imaging data which is critical for assessing tumour characteristics. PET-derived metrics encompassing key parameters such as maximum standardised uptake value, metabolic tumour volume, and total lesion glycolysis, are instrumental in prognosticating therapeutic response. Nonetheless, these parameters may not comprehensively capture the intricacies of spatial tumour complexity which is intrinsically linked with cellular and molecular dynamics, including proliferation and necrosis4,5. Intratumoural heterogeneity (ITH) may have prognostic clinical importance, particularly in appraising resistance to therapeutic interventions. Various image features, such as texture analysis, have been proposed for the evaluation of ITH6. Therefore, there is a need to quantitatively discern image patterns that have clinical relevance, particularly within the domain of spatial tumoural complexity, in order to discover and develop valuable clinical indicators and models for precise prognostication.
Radiomics analysis has recently emerged as a sophisticated, integrative, data-driven approach for quantifying complex image patterns and constructing clinically valuable models using imaging data. Extending beyond the limits of visual interpretation and basic quantitative assessment7, this approach employs automated, high-throughput extraction and refinement techniques to analyse a vast spectrum of image patterns from medical images. FDG- PET radiomics represents a noteworthy advancement in the quantitative analysis of imaging data and has attracted attention for its potential in assessing spatial tumoural complexity, including ITH, and predicting therapeutic response in the field of oncology. In this study, we aimed to develop and validate a model based on radiomics analysis of FDG-PET/computed tomography (CT) images of primary tumours to predict the overall survival (OS) of patients newly diagnosed with PDAC.
Patients and Overall Study Design
This study received institutional review board approval from Seoul Metropolitan Government - Seoul National University, Boramae Medical Center (SMG-SNU BMC; approval number: 30-2022-72), and was performed in accordance with the Declaration of Helsinki. The board granted a waiver of informed consent for this study due to its retrospective design, recognising its minimal risk to participants.
This retrospective, single-centre, medical record-review study was designed to predict the therapeutic response of patients who underwent FDG-PET/CT for initial staging of PDAC between June 2010 and October 2020. Patients were included if they met the following criteria: (a) the presence of FDG-avid primary tumour; (b) the absence of clinically suspected distant metastases based on CT, MRI, and FDG-PET/CT images; (c) the absence of any other primary malignancy; and (d) no history of receiving any anti-tumour treatment prior to PET/CT imaging, including surgery, chemotherapy, or radiotherapy. Patients were excluded if they had at least one of the following criteria: (a) tumour size too small to evaluate FDG distribution (< 64 voxels); (b) PET/CT misregistration due to respiratory motion; and (c) limited delineation of tumour uptake on FDG-PET/CT images due to adjacent activity of the stomach or inflamed pancreas. A flow chart depicting patient selection is shown in Fig. 1.
Prognostic models were constructed using both clinical parameters and FDG-PET/CT radiomics analysis, which were then validated to predict OS. The commencement of observation was defined as the first date of referral to SNM-SNU BMC for the management of pancreatic cancer. Each patient’s survival data was obtained from the national database of the Ministry of the Interior and Safety of South Korea, with December 2020 as the cut-off date. Demographic and clinical characteristic data were extracted from the medical records of each patient from the point of initial PDAC diagnosis.
FDG-PET/CT Imaging
Fasting status was confirmed by blood glucose levels (< 140 mg/dL). Subsequently, an intravenous injection of fluorodeoxyglucose (FDG) was administered at a dose of 5.18 MBq/kg. After a 60-minute rest period, a non-contrast CT scan was conducted using parameters of 80 mA and 140 kVp with a slice thickness of 5 mm. This was immediately followed by 1 minute of PET imaging per bed position utilising a dedicated PET/CT scanner (Gemini TF, Philips Healthcare, Cleveland, OH). CT images were reconstructed using a 512 × 512 matrix across a 50 cm field-of-view. PET scans covered the area extending from the mid-thigh to the vertex and were reconstructed using a 128 × 128 matrix. PET images were processed using the ordered subset expectation maximisation algorithm, with parameters of n subsets and m iterations, enhanced with time-of-flight and point spread function. Additionally, CT-based attenuation correction and application of a Gaussian filter with a full width at half maximum of 8 mm were employed to optimise image clarity and diagnostic accuracy.
Feature Extraction from FDG-PET/CT
A nuclear medicine physician (Y.K.K) with 10 years of experience performed PET/CT interpretation analysed PET/CT images using free medical image texture analysis software (LIFEx 7.4.0, http://www.lifexsoft.org ) 8. For segmentation of pancreatic tumours, spherical volumes of interest were manually delineated on PET images encompassing the entirety of each tumour, while a contrast-based approach was used to define tumour contours. Normal retroperitoneal fat tissue was excluded from the defined tumour volume via visual assessment of corresponding CT images. Texture parameters were automatically calculated from the final segmented tumour volume derived from PET and CT images. Intensity discretisation was carried out using a fixed bin size method for both PET (bin size, 0.5 standardized uptake value; range, 0–50) and CT (bin size, 10 Hounsfield unit; range, -150–150) images. A total of 222 texture parameters were used as input factors of the radiomics model. The extracted features were grouped under the following heading: including morphology, intensity, local intensity, intensity histogram, grey level co-occurrence, grey level run length, grey level size zone, and neighbourhood grey level difference features (Table S1).
Development of a Radiomics Model
A radiomics model was developed to predict OS. To identify parameters with significant prognostic influence among the extracted texture features, we employed the least absolute shrinkage and selection operator (LASSO) method for variable selection and shrinkage within the Cox proportional hazards regression analysis. We conducted 10-fold cross-validation on the entire patient cohort to minimize overfitting and optimise model hyperparameters. The model was optimised to maximise the Harrell’s concordance index (C-index). A radiomics score (Rad-score) was then produced for each patient by summing the selected prognostic variables weighted by their coefficients obtained from the LASSO Cox regression analysis. The Rad-score was used for the subsequent survival analysis.
Statistical Analysis
Clinical characteristics were compared between patients who survived and those who did not using Student’s t-test for continuous variables and chi-square test for categorical variables. The prognostic significance of clinical variables and the Rad-score for OS were assessed by univariate and multivariate Cox proportional hazards regression analyses, respectively. Harrell's C-index was used to represent the prognostic values of Cox models based on clinical variables (clinical model), Rad-score (Rad-score model), and a combination of both factors (clinical + Rad-score model). The difference in Harrell's C-index between the clinical model and the clinical + Rad-score model was assessed using bootstrapping with 1,000 resamples to determine the additive value of the Rad-score. Radiomics model development and statistical analyses were performed using R software 4.3.3 (http://www.R-project.org). LASSO Cox regression analyses were performed using the "glmnet" package embedded in R software (https://cran.r-project.org/web/packages/glmnet). A two-sided P-value below 0.05 was considered statistically significant. In the multivariate Cox analyses, variables were included in the model if their p-value was less than 0.05 and were removed if their p-value exceeded 0.1.
Patient Demographics and Clinical Characteristics
The study included 84 male and 56 female patients diagnosed with PDAC, with an average age of 69.0 ± 10.6 years. The clinical staging distribution was as follows: T1 stage, 10.7% (n = 15); T2, 27.9% (n = 39); T3, 25.0% (n = 35); and T4, 36.5% (n = 51). A considerable majority (75.0%) of patients had no suspected regional lymph node metastases. Most patients (78.6%; n = 110) received initial treatment at SNM-SNU BMC, including upfront surgery (36.36%; n = 40), systemic chemotherapy (30.91%; n = 34), and best supportive care alone (32.73%; n = 36). The Ministry of the Interior and Safety of South Korea’s national statistical database indicated that, among the included patients, 102 patients (72.8%) were deceased at the cut-off timepoint, with an average follow-up duration of 19.5 ± 19.2 months for the entire cohort, and 34.4 ± 28.8 months for survivors. No significant differences in demographic or clinical characteristics, including age, sex, smoking history, and clinical staging, were noted between survivors and non-survivors (Table 1).
| Factor | Survived (N = 38) | Deceased (N = 102) | Total (N = 140) | P* | |
|---|---|---|---|---|---|
| Age (year) | 67.3 ± 9.4 | 69.7 ± 11.0 | 69.0 ± 10.6 | 0.241 | |
| Sex (Male: Female) | 22:16 | 62:40 | 84:56 | 0.847 | |
| Follow-up periods (months) | 34.4 ± 28.8 | 14.0 ± 9.4 | 19.5 ± 19.2 | ||
| Smoking history | Smoker or ex-smoker | 5 (13.2%) | 25 (24.5%) | 30 | 0.147 |
| Never-smoker | 33 (86.8%) | 77 (75.5%) | 110 | ||
| Clinical T stage | T1 | 4 (10.5%) | 11 (10.8%) | 15 (10.7%) | 0.133 |
| T2 | 16 (42.1%) | 23 (22.5%) | 39 (27.9%) | ||
| T3 | 8 (21.1%) | 27 (26.5%) | 35 (25.0%) | ||
| T4 | 10 (26.3%) | 41 (40.2%) | 51 (36.5%) | ||
| Clinical N stage | N0 | 26 (68.4%) | 79 (77.5%) | 105 (75.0%) | 0.158 |
| N1 | 11 (28.9%) | 16 (15.7%) | 27 (19.3%) | ||
| N2 | 1 (2.6%) | 7 (6.9%) | 8 (5.7%) | ||
| Haemoglobin (g/dL) | 13.1 ± 1.8 | 12.5 ± 2.0 | 12.7 ± 1.9 | 0.096 | |
| WBC (109/L) | 7.1 ± 3.0 | 7.2 ± 2.9 | 7.2 ± 2.9 | 0.786 | |
| Platelet (103/µL) | 243.2 ± 62.5 | 260.0 ± 85.0 | 255.2 ± 79.6 | 0.277 | |
| AST (IU/L) | 148.5 ± 224.8 | 122.8 ± 149.3 | 129.7 ± 172.6 | 0.435 | |
| ALT (IU/L) | 166.2 ± 202.9 | 134.4 ± 171.2 | 143.1 ± 180.5 | 0.357 | |
| ALP (IU/L) | 303.9 ± 224.6 | 388.9 ± 387.9 | 365.9 ± 352.5 | 0.111 | |
| Total bilirubin (mg/dL) | 8.7 ± 9.4 | 7.5 ± 9.5 | 7.8 ± 9.5 | 0.491 | |
| CEA (ng/mL) | 3.6 ± 4.5 | 8.5 ± 23.6 | 7.1 ± 20.2 | 0.052 | |
| CA19-9 (U/mL) | 2,085.8 ± 7,151.4 | 2,427.6 ± 12,330.4 | 2,333.5 ± 11,122.2 | 0.873 | |
| *Student’s t tests for continuous parameters and chi square tests for categorical parameters were performed. | |||||
| WBC, white blood cell; AST, aspartate transaminase; ALT, alanine transferase; ALP, alkaline phosphatase; CEA, carcinoembryonic antigen; CA19-9, carbohydrate antigen 19 − 9 | |||||
Radiomics Model Development
LASSO Cox proportional hazards regression model assisted with optimal hyperparameters selected seven features of PET images and four features of CT images from among the 222 texture features extracted from PET/CT images (Figure S1 in Supplementary data) to construct the radiomics model. Details of the selected features and Rad-score calculation are as follows:
where MORPH_AV represents Surface To Volume Ratio, IH_SKEW represents Intensity Histogram Skewness, IH_P10 represents Intensity Histogram 10th percentile, IH_energy represents Intensity Histogram Energy, GLCM_NIDM represents GLCM Normalized Inverse Difference Moment, NGTDM_STR represents NGTDM strength, GLSZM_LZLGE represents GLSZM Large Zone Low Grey Level Emphasis, SKEW represents Intensity-based Skewness, MIN represents Minimum Grey Level, GLRLM_LRLGE represents GLRLM Long Run Low Grey Level Emphasis, and GLSZM_GLNU_NORM represents GLSZM Normalized Grey Level Non-uniformity. 'PET_' and 'CT_' prefixes indicate factors derived from PET and CT images, respectively. The composite Rad-score demonstrated significant prognostic value with a C-index of 0.671.
Prognostic Significance of Clinical Variables and Radiomics Score
Univariate Cox proportional hazards regression analysis to assess the prognostic value of clinical parameters and the Rad-score (Table 2) indicated that significant predictors of OS in patients with PDAC were age (HR, 1.040; 95% CI, 1.016–1.064; P = 0.0027), serum haemoglobin (Hb) level (HR, 0.840; 95% CI, 0.750–0.94; P < 0.0001), alkaline phosphatase (ALP) level (HR, 1. 001; 95% CI, 1.000–1.001; P = 0.0139), and carcinoembryonic antigen (CEA) level (HR, 1.021; 95% CI, 1.012–1.031; P < 0.0001). Smoking history showed borderline significance (HR, 1.510; 95% CI, 0.942–2.420; P = 0.0869) in predicting survival. T stage (4 vs. <4), N stage, and other laboratory test results had no significant prognostic impact. Notably, the Rad-score outperformed all clinical parameters with a higher hazard ratio (HR,12.698; 95% CI, 5.743–28.074; P < 0.0001). In Kaplan-Meier curve analysis with log-rank test using an optimal cutoff that produced the highest chi-square of Log-rank test, patients with higher and lower Rad-score demonstrated significant differences in overall survival (median OS 10.8 [95% CI 8.4–14.2] vs. 27.1 [19.6–34.1] months; chi-square = 29.254; P < 0.001; Fig. 2).
| Factor | Hazard ratio [95% CI] | P |
|---|---|---|
| Age | 1.040 [1.016–1.064] | 0.001 |
| Sex | 1.022 [0.678–1.541] | 0.918 |
| Smoking history | 1.510 [0.942–2.420] | 0.087 |
| Alcohol intake | 1.021 [0.673–1.549] | 0.922 |
| T4 (vs. T1-3) | 1.283 [0.850–1.937] | 0.235 |
| N+ | 0.818 [0.508–1.319] | 0.411 |
| WBC | 1.070 [0.991–1.154] | 0.082 |
| Haemoglobin | 0.840 [0.750–0.941] | 0.003 |
| Platelet | 1.002 [0.999–1.004] | 0.207 |
| AST | 1.000 [0.999–1.001] | 0.551 |
| ALT | 1.000 [0.999–1.001] | 0.774 |
| ALP | 1.001 [1.000–1.001] | 0.014 |
| Total bilirubin | 1.010 [0.988–1.033] | 0.369 |
| CEA | 1.021 [1.012–1.031] | < 0.001 |
| CA19-9 | 1.000 [1.000–1.000] | 0.175 |
| Rad-score | 12.698 [5.743–28.074] | < 0.001 |
N+, presence of regional lymph node metastasis; WBC, white blood cell; AST, aspartate transaminase; ALT, alanine transferase; ALP, alkaline phosphatase; CEA, carcinoembryonic antigen; CA19-9, carbohydrate antigen 19-9
Additive Value of the Radiomics Score to Clinical Prognostics
Multivariate Cox proportional hazards regression analysis compared the prognostic value of clinical and combined (clinical + Rad-score) survival prediction models (Table 3). In the clinical model, age, smoking history, serum ALP level, CEA level, and T stage were identified as independent predictors of OS. In the clinical + Rad-score model, Rad-score was as a significant prognostic factor (HR, 12.793; 95% CI, 5.54–29.546; P < 0.0001). Thus, Rad-score was included in the model and T stage was excluded. Inclusion of Rad-score in the clinical model improved prognostic accuracy as evidenced by a higher Harrell's C-index (0.740 vs. 0.673). Bootstrapping analysis showed that the difference in prognostic value between the clinical and clinical + Rad-score models was statistically significant (difference in C-index, 0.069; 95% CI, 0.014–0.113; P < 0.0001).
| Factor | Clinical model | Combined model | ||
|---|---|---|---|---|
| Hazard ratio [95% CI] | P | Hazard ratio [95% CI] | P | |
| Age | 1.050 [1.025–1.077] | < 0.001 | 1.052 [1.026–1.078] | < 0.001 |
| Smoking history | 2.067 [1.242–3.440] | 0.005 | 2.080 [1.237–3.498] | 0.006 |
| T4 (vs. T1-3) | 1.454 [0.940–2.251] | 0.093 | N/S | |
| N+ | N/S | N/S | ||
| Haemoglobin | N/S | N/S | ||
| ALP | 1.001 [1.000–1.001] | 0.043 | 1.001 [1.000–1.001] | 0.034 |
| CEA | 1.021 [1.011–1.031] | < 0.001 | 1.014 [1.004–1.025] | 0.007 |
| CA19-9 | N/S | N/S | ||
| Rad-score | - | - | 12.793 [5.540–29.546] | < 0.001 |
| C-index | 0.673 [0.650–0.766] | 0.740 [0.715–0.816] | ||
| C-index difference* | 0.067 [0.014–0.113] | |||
| *Difference between two models produced by bootstrapping 1,000 resamples. | ||||
| N/S, not significant; ALP, alkaline phosphatase; CEA, carcinoembryonic antigen; CA19-9, carbohydrate antigen 19 − 9 | ||||
We successfully formulated a prognostic model for the evaluation of patients with PDAC using FDG-PET/CT-based radiomics. By identifying key features through texture analysis of FDG-PET/CT images and integrating them into a composite Rad-score, we constructed a model with a nuanced approach to oncologic prognostication. The Rad-score was validated, demonstrating statistically significant, albeit moderate, prognostic ability to discriminate between survivors and decedents, indicating its potential utility in clinical prognostication. A combined model, incorporating both clinical parameters and Rad-score significantly outperformed the use of clinical parameters alone for predicting OS in patients with PDAC, indicating that Rad-score may be a crucial biomarker for survival outcomes. In addition, multivariate Cox proportional hazards regression analysis demonstrated that integrating the Rad-score with clinical variables significantly enhanced the prognostic accuracy of OS prediction models. This study underscores the critical contribution of radiomics in refining the accuracy of survival prediction and its potential in guiding personalised treatment in oncology.
Several studies have demonstrated the prognostic value of radiomics in pancreatic cancer, mainly based on contrast-enhanced CT or MRI9. However, research focusing on FDG PET-based radiomics for predicting OS in pancreatic cancer remains limited. Toyama et al. showed the prognostic value of GLZLM grey-level non-uniformity in FDG-PET using a random forest classifier, but this study only estimated survival outcomes over a one-year observation period10. Hyun et al. analysed a cohort of 137 PDAC patients and reported that first-order entropy on initial FDG-PET independently correlates with survival, as determined through multivariate Cox regression analysis11. Similarly, JW Lee et al. demonstrated the prognostic significance of first-order entropy and developed and reported a scoring system that incorporates total lesion glycolysis and bone marrow uptake to predict OS12. Conversely, Yoo et al., in a study of patients with pancreatic cancer undergoing curative surgery, failed to identify independent predictive factors from heterogeneous features measured on FDG-PET through multivariate analysis13. Despite providing us with valuable insights, the previous studies are somewhat limited by their focus on individual radiomics features. These studies used conventional Cox regression analysis, which has inherent shortcomings such multicollinearity, ineffective variable selection, and potential overfitting. Addressing these issues, we employed LASSO for robust variable selection and designed an integrated prognosis model to emphasize the power of integrated radiomics data. Nonetheless, we highlight the need for ongoing optimisation to improve the predictive performance of the model. Additionally, we investigated the additive value of radiomics data by comparing it with a prognosis model based on clinical variables alone, further validating our approach through internal bootstrapping analysis.
In our study, the Rad-score emerged as an independent prognostic indicator in the multivariate analysis, while T stage did not retain its prognostic significance within the radiomics-enhanced model (clinical + Rad-score model). This change suggests the potential of the radiomics model to serve as a more comprehensive and robust prognostic tool by capturing the intricate characteristics of primary tumors, which may not be fully appreciated through traditional clinical markers alone. Additionally, while certain clinical prognostic factors such as age, smoking history, and carcinoembryonic antigen (CEA) levels aligned with findings from previous studies, other factors like N stage and carbohydrate antigen 19 − 9 (CA19-9) did not14–18. This discrepancy may be attributed to the heterogeneous disease status and the variability in treatment and follow-up strategies among patients involved in the study. The need for further validation in a large multi-center cohort with unified management strategies remains to solidify its usefullness in clinical practice.
This study has several limitations. First, its retrospective design has an inherent risk of bias because it includes patients who received different treatment regimens, several of whom received follow-up treatment after initial diagnosis at hospitals other than SNM-SNU BMC. However, given the strict regulation of cancer treatment by the Korean national insurance system, it is plausible that most patients received standardised treatment in accordance with national cancer treatment guidelines (references). Second, while this study included a larger number of patients compared with previous studies, its single-centre design may limit the generalizability of the results. Prospective multi-centre studies with external data are warranted. Third, lesions were excluded from analysis if there was insufficient tumour size for evaluating FDG distribution, PET-CT misregistration due to respiratory motion, or challenges in delineating tumour uptake on FDG-PET/CT images. This exclusion criteria are a common challenge in PET radiomics analyses, because it safeguards appropriate texture analysis6.
An FDG-PET/CT-based radiomics model showed potential in enhancing the prediction of survival outcomes among patients with PDAC, and outperformed a model based on clinical data alone demonstrating its potential applicability in the field. Further prospective studies with larger cohorts are warranted to validate the results of the current study and establish the model’s applicability in patient management.
Acknowledgements
The authors declare no conflict of interest.
Competing Interests Statement
The authors declare no competing interests.
Author Contribution
Y.K., S.H., J.B.J., and S.W.O. designed and conceptualized the study. Y.K. and S.H. collected data, and performed radiomics analysis and statistical analysis. J.B.J. and S.W.O. critically discussed the analysis and results. Y.K. and S.H. initially wrote the manuscript. J.B.J. and S.W.O. revised the manuscript. All these authors read and approved the final version of the manuscript.
Data Availability
The datasets generated or analyzed during the study are available from the corresponding author on reasonable request. Due to patient confidentiality agreements, raw imaging data are not publicly available but can be accessed upon approval by the institutional review board.
- Rawla, P., Sunkara, T. & Gaduputi, V. Epidemiology of Pancreatic Cancer: Global Trends, Etiology and Risk Factors. World J Oncol 10, 10–27, doi:10.14740/wjon1166 (2019).
- De Dosso, S. et al. Treatment landscape of metastatic pancreatic cancer. Cancer Treat Rev 96, 102180, doi:10.1016/j.ctrv.2021.102180 (2021).
- Bengtsson, A., Andersson, R. & Ansari, D. The actual 5-year survivors of pancreatic ductal adenocarcinoma based on real-world data. Sci Rep 10, 16425, doi:10.1038/s41598-020-73525-y (2020).
- Chicklore, S. et al. Quantifying tumour heterogeneity in 18F-FDG PET/CT imaging by texture analysis. Eur J Nucl Med Mol Imaging 40, 133–140, doi:10.1007/s00259-012-2247-0 (2013).
- Gillies, R. J., Kinahan, P. E. & Hricak, H. Radiomics: Images Are More than Pictures, They Are Data. Radiology 278, 563–577, doi:10.1148/radiol.2015151169 (2016).
- Ha, S., Choi, H., Paeng, J. C. & Cheon, G. J. Radiomics in Oncological PET/CT: a Methodological Overview. Nucl Med Mol Imaging 53, 14–29, doi:10.1007/s13139-019-00571-4 (2019).
- Hatt, M. et al. Characterization of PET/CT images using texture analysis: the past, the present… any future? Eur J Nucl Med Mol Imaging 44, 151–165, doi:10.1007/s00259-016-3427-0 (2017).
- Nioche, C. et al. LIFEx: A Freeware for Radiomic Feature Calculation in Multimodality Imaging to Accelerate Advances in the Characterization of Tumor Heterogeneity. Cancer Research 78, 4786–4789, doi:10.1158/0008-5472.Can-18-0125 (2018).
- Gao, Y. et al. A systematic review of prognosis predictive role of radiomics in pancreatic cancer: heterogeneity markers or statistical tricks? European Radiology 32, 8443–8452, doi:10.1007/s00330-022-08922-0 (2022).
- Toyama, Y. et al. Prognostic value of FDG-PET radiomics with machine learning in pancreatic cancer. Scientific Reports 10, 17024, doi:10.1038/s41598-020-73237-3 (2020).
- Hyun, S. H. et al. Intratumoral heterogeneity of 18F-FDG uptake predicts survival in patients with pancreatic ductal adenocarcinoma. European Journal of Nuclear Medicine and Molecular Imaging 43, 1461–1468, doi:10.1007/s00259-016-3316-6 (2016).
- Lee, J. W., Park, S.-H., Ahn, H., Lee, S. M. & Jang, S. J. Predicting Survival in Patients with Pancreatic Cancer by Integrating Bone Marrow FDG Uptake and Radiomic Features of Primary Tumor in PET/CT. Cancers 13, 3563 (2021).
- Yoo, M. Y. et al. Prognosis prediction of pancreatic cancer after curative intent surgery using imaging parameters derived from F-18 fluorodeoxyglucose positron emission tomography/computed tomography. Medicine 99, e21829, doi:10.1097/md.0000000000021829 (2020).
- Allen, P. J. et al. Multi-institutional Validation Study of the American Joint Commission on Cancer (8th Edition) Changes for T and N Staging in Patients With Pancreatic Adenocarcinoma. Annals of Surgery 265, 185–191, doi:10.1097/sla.0000000000001763 (2017).
- Dell'Aquila, E. et al. Prognostic and predictive factors in pancreatic cancer. Oncotarget 11, 924–941, doi:10.18632/oncotarget.27518 (2020).
- Ueda, M. et al. Prognostic Factors After Resection of Pancreatic Cancer. World Journal of Surgery 33, 104–110, doi:10.1007/s00268-008-9807-2 (2009).
- Ni, X. G. et al. The clinical value of serum CEA, CA19-9, and CA242 in the diagnosis and prognosis of pancreatic cancer. European Journal of Surgical Oncology (EJSO) 31, 164–169, doi:https://doi.org/10.1016/j.ejso.2004.09.007 (2005).
- Zell, J. A., Rhee, J. M., Ziogas, A., Lipkin, S. M. & Anton-Culver, H. Race, Socioeconomic Status, Treatment, and Survival Time among Pancreatic Cancer Cases in California. Cancer Epidemiology, Biomarkers & Prevention 16, 546–552, doi:10.1158/1055-9965.Epi-06-0893 (2007).
No competing interests reported.