Joint Radiomics Analysis on 18F-FDG-PET-MRI of Primary Tumor to Predict the Risk of Synchronous Distant Metastasis in Pancreatic Ductal Adenocarcinoma

Purpose: This study explores the potential of joint radiomics analysis of PET-MRI of primary pancreatic ductal adenocarcinoma (PDAC) tumor in predicting the risk of synchronous distant metastasis (SDM). Methods: Two cohorts of PDAC patients were collected including a development cohort (n=66) receiving separate 18 F-FDG-PET/CT and multi-sequence MRI, and a external test cohort (n=25) receiving hybrid PET/MR. All of these patients were confirmed with SDM after imaging. Radiomics features of primary PDAC tumors were selected and models were built for PET, MRI, and PET-MRI from the development cohort. A radiomics nomogram was constructed by combining independent clinical indicators. The developed radiomics nomogram was independently evaluated on the test cohort. Results ： The area under the curve (AUC) values of PET, MRI, and joint PET-MRI models were 0.89, 0.86, and 0.94 in the training set and 0.77, 0.67, and 0.77 in the test set. The radiomics nomogram combing the joint PET-MRI radiomics signature, age, and CA19-9 level had good calibration and high discrimination capacity with maximum AUC value (0.93). The decision curve analysis (DCA) confirmed the radiomics nomogram had clinical usefulness. The evaluation on the independent test cohort showed that the accuracy, sensitivity, specificityand AUC values of radiomics nomogram were 84.0%, 78.6%, 90.9% and 0.85. Conclusion The robust and effective prediction of the risk of SDM for the preoperative PDAC patients confirmed the potential of the radiomics analysis on PET/MR. The radiomics information in primary tumor may provide complementary and alerting hints for cancer staging.


Introduction
Pancreatic ductal adenocarcinoma (PDAC)is characterized by late diagnosis, high mortality rate, and poor prognosis [1,2]. Approximately 50% of new PDAC cases were detected with distant metastases at diagnosis of the primary tumor [3], which would preclude a radical therapeutic approach and indicate poor prognosis [4]. And most patients undergoing surgical resection will have metastasis within 4 years after surgery, which indicates that micrometastasis actually exists in patients with obviously limited tumors [5]. The treatment options available for patients with pancreatic cancer depend on the primary disease stage. Patients are eligible for curative surgery for nonmetastatic tumors and metastatic tumors may benefit from palliative chemotherapy. PDAC is characterized by a higher heterogeneous risk of metastasis [6]. Therefore, the detection of distant metastases is imperative for sparing from unnecessary pancreatectomy and selecting appropriate treatment strategies in PDAC patients.In addition, neoadjuvant treatment is the accepted approach for resectable patients with high risk features according to 2021 NCCN guidelines [7]. 18 F-fluorodeoxyglucose (FDG) positron emission tomography/computed tomography (PET/CT) is a diagnostic tool in detecting distant metastases [8,9].However, it is suboptimal in the evaluation of liver lesions due to high background hepatic FDG uptake [10] and the detection capability of subtle peritoneal carcinomatosis in PDAC is still limited [11]. Compared with CT, multi-phase contrastenhanced (CE) sequences and diffusion-weighted imaging (DWI) sequences of MRI are useful for the presence of liver metastases [12,13] and peritoneal metastases [14] from PDAC. Moreover, FDG-PET and MRI can provide both functional metabolic information from PET and better tissue contrast and anatomical resolution from MRI.
However, FDG-PET might miss liver/peritoneal metastatic tumors that were <1 cm [15] and MRI had limitations in detecting liver metastasis smaller than 3mm [16] and lung metastasis. Therefore, a new tool that could improve the sensitivity and accuracy of imaging techniques are needed.
Radiomics converted digital medical images into high-throughput quantitative features for clinical decision support and improved diagnostic accuracy [17]. Tumor intensity, shape, and textural features extracted from radiomics are thought to have the potential to reflect intratumoral histopathological properties [18].A recent study demonstrated the use of nomograms based on clinicopathological features for predicting liver metastasis in 12644 patients with PDAC [19].However,pathologic information was obtained by examining the surgical specimen and thus cannot be applied to guide therapeutic strategies. Accordingly, preoperative noninvasive biomarkers are needed to predict the occurrence of synchronous distant metastasis (SDM). 18 F-FDG-PET and CE-MRI offer the opportunity to study tumour pathophysiology [20]. A recent study was performed in 17 patients with PDAC by Joo et al and the sensitivity of FDG PET/MR to assess M stage was 75% preoperatively [21].
However, there are few studies that used PET/MR radiomics to predict the risk of SDM in pancreatic cancer patients [22]. In addition, features extracted from multiparametric methods including T2WI, DWI, and CE-MRI sequences have lower risk of bias than from one sequence alone [23]. We hypothesize that radiomics signature using features derived from 18

PET/MR acquisition
Whole-body PET/MR was performed using a Siemens Biograph mMR scanner (Siemens Healthcare, Erlangen, Germany). All participants were fasted for atleast 6hours given intravenous 18 F-FDG 2.5 to 6MBq/kg 40-100minutes before the study.

Segmentation and Radiomic features extraction
Three-dimensional volumes of interest (VOIs) were manually and independently drawn by two nuclear medicine physicians on 18 F-FDG-PET/CT and MRI. Image features were extracted using LIFEx plat form v5.1 for radiomics [25]. All images underwent standardized preprocessing as follows: intensity discretization to 64 gray levels,spatial resampling was set to 2.0 × 2.0 × 2.0 mm 3 .Absolute intensity rescaling using 0-20 SUV units in the 18

Construction of the Radiomics Models, Clinical Model and Radiomics Nomogram
To prevent overfitting in radiomics models, feature subset selection of highdimensional image data was conducted using the following steps. Features with ICCs > 0.75 were retained for subsequent analysis across all sequences to avoid interobserver variability in segmenting the VOIs.The least absolute shrinkage and selection operator (LASSO) logistic regression [27] method was performed to select metastasis-related features of high-dimensional data [28]and to build radiomics model. The penalty parameter tuning was conducted with the smallest 10-fold cross-validation error.
Significant features with nonzero coefficients were selected. A radiomics model was a linear combination of selected features weighted by their respective coefficients. We developed three radiomics models: two were based on features extracted from PET or MRI images, and the third was a combination of both PET and MRI texture parameters (PET-MRI). The features and radiomics models were generated from the training set, and then the testing set was used to confirm our findings in independent cases. The best radiomics signature and clinic-pathological information were used to construct multivariate regression in the training set. The backward stepwise selection was applied by using the likelihood ratio test with Akaike's Information Criterion (AIC) as the stopping rule [29]. Varabilaities with P < 0.05 in multivariate logistic regression were introduced to build a radiomics nomogram. The clinical model was based on nomogram except for a radiomics signature.

Statistical analyses
R software (version 3.6.1, http://www.r-project.org)and SPSS 25.0 (IBM, Armonk, NY) were used for performing statistical analysis. P< 0.05 (two-sided) was considered statistically significant. Independent sample t-tests and Fisher's exact tests were used for comparing continuous variables and categorical variables between the two groups, respectively. The area under the curve (AUC) of the receiver operator characteristic (ROC) curve was used to evaluate the discrimination performance of the radiomics models in both the training and testing sets. The diagnostic abilities between each two of the three models were compared by using the Delong test [30]. The Hosmer-Lemeshow test was used to assess the goodness-of-fit of the radiomics nomogram.
Decision curve analysis (DCA) was used to quantify the net benefit of the models at different threshold probabilities [31].Spearman's correlation coefficient (ρ) was calculated to examine the correlation between radiomics features and clinicpathological characteristics. The accuracy, sensitivity and specificity were computed by Confusion matrix-derived metrics.The flow chart of patient recruitment is shown in Fig.   1.

Patient clinical characteristics
The clinical characteristics of the 91 patients in the development and test cohorts are summarized in Table 1

Feature Selection and Radiomics Model Building
We used the LASSO algorithm to select the optimal subset of radiomics features in the training set. Six radiomics features with nonzero coefficients were selected using PET-MRI images (Fig. 3). Four and Five radiomics features were selected using PET images and MRI images, respectively. The three radiomics models constructed using optimal features weighted respective coefficients by the LASSO regression from each group were as follows：

Performance of different radiomics Model
The ROCs of all three abovementioned radiomics models are provided in Fig. 3Aand 3B  Table 2 shows the diagnostic performance for three radiomics models. The radiomics signature of the PET-MRI with the best performance was being applied to the subsequent analysis.

Development and evaluation of the Radiomic Nomogram
In the multivariate analysis (Table 3), CA19-9 level, age and the radiomics signature were significantly associated with SDM status. A radiomics nomogram was then constructed with radiomics signature, the CA19-9 level and age (Fig. 5) Fig 1).

Prediction of SDM from primary tumor
PDAC is a highly invasive tumor and often leads to distant metastasis. The SDM of pancreatic cancer may be related to the gene changes, tumor microenvironment and other factors. Several altered genes in PDAC, such as, KRAS (90%), TP53, CDKN2A, and SMAD4 [32,33], were identified to be associated with the invasion and metastasis.
Radiomics can unravel the heterogeneity of the tumor [34], which is associated with underlying changes of genetics and tumor microenvironment [35]. The morphological heterogeneity of primary tumors is considered as an important determinant factor of metastatic phenotype [36].

Complementary information between PET and MRI
PET and MRI decode different (patho-)physiological mechanisms that complement each other and may be accessible through textural feature analysis [38]. A combined, multimodal predictor has the potential to outperform single modality models [39]. This assumption is supported by our findings where the combination of both modalities yielded the highest diagnostic performance compared to the single modality models.
PET and MRI are routine powerful diagnostic tools. The sensitivity of the radiomics signature reached 93.9% in the training set. Compared with the imaging report, our study found that radiomics signature derived from PET and MRI imaging increase the diagnostic sensitivity of SDM in PDAC (65% for 18 F-FDG PET/CT, and 80% for MRI).In addition, our results from the tumor radiomics setup show that different sequences (i.e., PET, T2, ADC and T1WI in arterial phase) can provide metabolic, anatomic, and functional information. ADC is a quantitative biomarker calculated from diffusion-weighted imaging that can reflect tumor characteristics such as cellular density and aggressiveness [40].Choi et al demonstrated that entropy was significantly associated with overall survival in patients with PDAC using preoperative T2WI texture analysis [41].Previous studies have found integrated PET/MR to be able to assess treatment response [42], prognosis [43] in PDAC patients.

Interpretation of the identified radiomics features
Radiomics can provide comprehensive tissue and organ characterization undetected by human perception and is a fast, low-cost and noninvasive approach. Our optimal radiomics signature contained 3 features on 18 F-FDG PET images and 3 features on MRI. As a conventional metabolic feature,SUVpeak means maximum average SUV in a 1 cm 3 sphere of the VOI, which can reflect the metabolic activity. Higher pretreatment SUVpeak in patients with advanced pancreatic cancer indicated a poor prognosis. Kim et al. showed that the SUVpeak was an independent prognostic factor in lung adenocarcinoma [44].The shape-based feature SHAPE_Volume (ml) from PET is the volume of interest. A previous study demonstrated that the probability of tumor dissemination was associated with tumor size at the time of diagnosis [45]. Regarding sphericity on T2WI, the lower the value, the further the divergence from a sphere.
Aggressive tumors may have more irregular boundaries and invasion of surrounding tissues on T2WI in patients with non-metastasis than in those with distant metastasis.
The remaining texture features, including PET texture features (GLZLM_SZE), ADC texture features (GLZLM_LZE), and AP texture features (GLZLM_SZE) reflect the heterogeneity and image uniformity within the volume of interest.

Value of clinical parameters
Our study confirmed that the high-performance radiomic nomogram included clinical factors (age and CA19-9 level) and the radiomic signature. Consistent with previous studies, CA-199 was identified as an independent predictor for distant metastasis in a previous study [46]. Dong et al found that the elevated level of CA19-9 was significantly associated with the risk of synchronous liver metastasis in stage IV PDAC patients [47]. Age > 62 years was correlated with the unresectability of synchronous liver metastasis in patients with PDAC [24]. However, the clinical model included CA19-9 and age achieved an AUC of 0.70 with limited value for the prediction of SDM in PDAC.

Robustness of the identified radiomics features
A critical challenge for the application of radiomics models is robust when applying to different scanners or imaging protocols.Generally, selecting reproducible features across different machines and protocols is important for the development and validation of radiomics models [48]. Often, standardization strategies are recommended to bridge the difference of voxel sizes, intensity outlier filtering, intensity discretisation and ICC, which is key to improve reproducibility of radiomic models [49]. In this study, radiomics nomogram was constructed on data from individual PET and MRI and then validated on hybrid PET/MRI scanner. Considering that the images were obtained in different magnet strength (1.5 T, 3.0 T) and with varying parameters, we applied feature normalization to reduce variations among patients.Even though we did not apply harmonization on the radiomics features [50], our current preliminary results demonstrated that certain robustness of the developed methods in cross-scanner applications. Additional data standardization and harmonization may further improve the performance in different scanner and maybe still necessary at extreme situation.
Nevertheless, this baseline cross-scanner robustness may be convenient for clinical translation, where the radiomics nomogram may be independently applied in many applications [51].

The application of nomogram
The predictive nomograms developed in this study may help clinicians quantitatively assess the probability of synchronous distant metastasis and make individualized decisions. For example, a 60-year patient with normal CA19-9 levels and low score radiomics signature is associated less than 10% probability of having synchronous metastasis, and might benefit from radical surgery. Neoadjuvant therapy are recommended for patients who are at high-risk of distant metastasis but the imaging are not able to reveal the lesion. The benefits of neoadjuvant therapy include eliminating micrometastases, and identifying aggressive tumors to avoid futile surgery [52]. At present, whether to carry out neoadjuvant treatment is decided by consulting the surgeon.Our nomogram will identify patients with high-risk metastatic who may benefit from a personalized approach for tumours that are susceptible to neoadjuvant treatment that could extend survival.

Limitations
This study has limitations. First, our data were retrospective, and a small number of patients were involved in this preliminary study due to a few patients receiving imaging examination, larger sample sizes and multicenter evaluation might be needed to validate our radiomics nomogram in the future.Second, The patients included in this study are involved in multiple organ metastases, which can not be used to predict the specific location of distant metastasis in patients with pancreatic ductal adenocarcinoma, but can be used to predict whether metastasis occurs, so as to help patients choose the optimal treatment. Finally, we used only one method for feature selection and classifier.
Further investigations may be performed to confirm whether our model be independent of other feature selection and classifier algorithms.

Conclusions
In summary, our study developed a noninvasive and quantitative tool for the

Funding information
This study was supported by research grants from the National Natural Science

Compliance with ethical standards
Conflict of interest All other authors have no conflicts of interest.

Research involving human participants
All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards.
Informed consent Informed consent was waived for this retrospective study.