A radiomics nomogram based on 18F-FDG PET/CT and clinical risk factors for the prediction of peritoneal metastasis in gastric cancer

Purpose Peritoneal metastasis (PM) is usually considered an incurable factor of gastric cancer (GC) and not fit for surgery. The aim of this study is to develop and validate an 18F-FDG PET/CT-derived radiomics model combining with clinical risk factors for predicting PM of GC. Method In this retrospective study, 410 GC patients (PM − = 281, PM + = 129) who underwent preoperative 18F-FDG PET/CT images from January 2015 to October 2021 were analyzed. The patients were randomly divided into a training cohort (n = 288) and a validation cohort (n = 122). The maximum relevance and minimum redundancy (mRMR) and the least shrinkage and selection operator method were applied to select feature. Multivariable logistic regression analysis was preformed to develop the predicting model. Discrimination, calibration, and clinical usefulness were used to evaluate the performance of the nomogram. Result Fourteen radiomics feature parameters were selected to construct radiomics model. The area under the curve (AUC) of the radiomics model were 0.86 [95% confidence interval (CI), 0.81–0.90] in the training cohort and 0.85 (95% CI, 0.78–0.92) in the validation cohort. After multivariable logistic regression, peritoneal effusion, mean standardized uptake value (SUVmean), carbohydrate antigen 125 (CA125) and radiomics signature showed statistically significant differences between different PM status patients(P < 0.05). They were chosen to construct the comprehensive predicting model which showed a performance with an AUC of 0.92 (95% CI, 0.89–0.95) in the training cohort and 0.92 (95% CI, 0.86–0.98) in the validation cohort, respectively. Conclusion The nomogram based on 18F-FDG PET/CT radiomics features and clinical risk factors can be potentially applied in individualized treatment strategy-making for GC patients before the surgery.


Introduction
Gastric cancer (GC) is the fourth leading cause of cancer-related death worldwide and the fifth commonest cancer worldwide [1].Gastrectomy is the main treatment for early-stage GC patients.The peritoneum is the most probable location of distant metastasis in GC [2].The 5-year survival rate of the patients with peritoneal metastasis (PM) is only 2% and the median survival time will decrease to only 3-5 months due to PM is an incurable factor of GC [3,4].It is unnecessary for a GC patient with PM to accept a surgical resection whose clinical benefit is unclear.So far, the effective treatment for PM is systemic chemotherapy or molecular targeted mechanisms [5,6].Thus, accurate preoperative prediction of PM in GC patients is vital for prognosis and treatment decisions.
Abdominal computed tomography (CT) is the primary diagnostic modality for pretreatment GC clinical staging.Several signs seen on CT imaging like omental cake, diffuse peritoneum thickening or massive peritoneal effusion could be considered existence of PM and usually in advanced state [7], but there is a problem that the CT-diagnosed PM-negative GC patients may confirmed positive by later peritoneal lavage cytology [8,9].These patients may miss the appropriate time window of systemic chemotherapy. 18F-FDG PET/CT is widely used in different kinds of malignant tumors diagnosis [10].It acquires imaging of functional information combined with anatomic data at a single device in one diagnostic session and provides a higher diagnostic accuracy than CT for detection of distant metastasis like PM [11,12].However, during the diagnosis of PM using 18 F-FDG PET/CT, there are still some defects which cannot be ignored.For example, due to the existence of intestinal physiological uptake or a small and low SUV lesion, the lesion may be easily missed in one diagnostic process.
Radiomics, an emerging field that involves great quantity of datasets, could extract and select features from medical imaging even if they are difficult to recognize and quantify from human eyes [13].The features provide additional information to predict underlying tumor biology and behavior [14].Moreover, these features can be used alone or with other patient-related data, and may potentially reveal the prognosis or clinical response to the new therapy [15][16][17].Recently, radiomics has been widely applied in gastrointestinal tumors for Lauren classification predicting, survival and chemotherapeutic benefits analysis, assessing GC HER2 expression [18][19][20].
Our previous research [21] has established a radiomics model which used a single modality based on 18 F-FDG PET and showed satisfied diagnostic ability in differentiating PM-positive GC patients, while anatomical information provided by CT images should not be neglected during the clinical practice.Hence, in this study, we aim to integrate advantages of 18 F-FDG PET and CT features and combine the clinical risk factors, to further explore the clinical capability of the radiomics nomogram to predict PM status in patients with GC.

Patient
This study was approved by the ethics committee of our hospital and the informed consent requirement was waived.All patients involved were collected from January 2015 to October 2021.The inclusion criteria of this study were as follows: (I) underwent 18 F-FDG PET/CT examination, (II) diagnosed GC by pathology or endoscopy, (III) peritoneal status confirmed by operation and pathology.The exclusion criteria were as follows: (I) combined or history with other malignant tumors, (II) receive anti-cancer therapy before the 18 F-FDG PET/CT scanning, (III) lack of complete clinical data, (IV) no visible lesions on 18 F-FDG PET/CT images.Fundamental clinicopathological data such as clinical characteristics (sex, age, smoking, alcohol), symptoms (abdominal pain, fever, vomit, weight loss, peritoneal effusion), serum tumor markers [cancer antigen 19-9 (CA19-9), cancer antigen 125 (CA125), carcinoembryonic antigen (CEA)] were obtained from medical records.The patients were randomly divided into training and validation cohort at a fixed ratio (7:3).The enrollment process was shown in Fig. 1.Enrollment process.

PET/CT imaging
All patients were required to fast for at least 6 h before the scanning and controlled their blood glucose levels below 110 ml/dl.The blood glucose level was monitored by finger stick immediately before the injection of 18 F-FDG.An intravenous injection of 18 F-FDG was performed to all patients by the dosage of 3.7 MBq/kg, then, images were acquired by a hybrid PET/CT scanner (GEMINI TF 64, Philips, Netherlands) within 60 min approximately.The CT scans were performed first using a slice thickness of 3.0 mm and reconstructed to a 512 × 512 matrix.Similarly, the PET scans were performed within 1.5 min in each bed, images were reconstructed to a 144 × 144 matrix, the field of view is 576 mm, slice thickness, and interval is 5 mm.

Tumor segmentation
Figure 2 showed the radiomics procedure of this study.The 18 F-FDG PET images and CT images were both independently read by two radiologists with good experience in PET/CT diagnosis.All the images were read in the digital imaging and communications in medicine protocol.The region of interest (ROI) was segmented semi-automatically using LIFEx software tools.The ROI of CT covered each slice of the primary lesion, fat, air, normal tissues and organs in adjacent areas were carefully avoided during the drawing of ROI.For the ROI of PET, we also chose to contain all parts of the visible lesion, maximum standardized uptake value (SUVmax) was automatically measured by LIFEx software [22], after that, a threshold of 40% of SUVmax was performed to reduce the ROI area, mean standardized uptake value (SUVmean), metabolic tumor volume and total lesion glycolysis were also automatically determined based on this area.

Radiomics features extraction and selection
Two groups of quantified texture features (69 features in CT and 69 features in PET) were extracted using the LIFEX software.The features included First order metrics: histogram, skewness, kurtosis, entropy, energy, sphericity and compacity.Second order metrics: graylevel co-occurrence matrix, gray-level zone length matrix (GLZLM), neighborhood gray-level dependence matrix (NGLDM), and gray-level run length matrix (GLRLM).
In order to ensure reproducibility and robustness of the following extracted features, we used intra-class correlation coefficients (ICCs) to assess the observer agreement of the feature extraction between the two operators.The ICC values >0.75 were retained to consider a good agreement [23].The features' ICC values were shown in Supplementary Table 1, Supplemental digital content 1, http://links.lww.com/NMC/A253.We used the maximum relevance and minimum redundancy (mRMR) and 10-fold cross-validated least absolute shrinkage and selection operator (LASSO) method to select radiomics features from primary features.The LASSO included a regular parameter λ, which determined the number of features.At first, mRMR was performed to eliminate the redundant and irrelevant features [24], then LASSO was conducted to choose the optimized subset of features.With the parameter regulated, LASSO differentiated features which do not relevant to PM by shrinking their coefficients to zero, those features with non-zero coefficients were remained.The radiomics score (Rad-score) was finally calculated by determining the product of the selected features' linear combinations and their respective coefficients.
In this study, mRMR and LASSO were performed by the 'mRMRe' and 'glmnet' package respectively on the R software (version 4.1.1).All features extracted from the LIFEx software were shown in Supplementary Table 2, Supplemental digital content 2, http://links.lww.com/NMC/A254.

Development of the prediction model
We used multivariable logistic regression analysis to develop a prediction model by combining the radiomics signature and clinical predictors with P values less than 0.05 in the univariable analysis.Mann-Whitney U test was used to compare the clinical predictors and PET parameters.To promote the clinical utility of the prediction model, we visualized the model as a radiomics nomogram based on multivariable logistic analysis.We used the area under the curve (AUC) of the receiver operating characteristic curve to assess the performance of the radiomics nomogram.Decision curve analysis was performed to evaluate the clinical usefulness of the radiomics nomogram by calculating the net benefits.

Statistical analysis
We used IBM SPSS Statistics (Version 22.0; IBM) to analyze the clinical data.The t test and Mann-Whitney U test were used to evaluate the differences in numerical variables such as age, CA125, CA19-9, CEA, and the Chi-square test or Fisher's exact test were used to analyze categorical variables, such as sex, alcohol, symptoms (abdominal pain, fever, vomit, weight loss, peritoneal effusion).R software (Version 4.1.1;R Foundation for Statistical Computing, Vienna, Austria, http://www.R-project.org) was used to analyze the radiomics feature data.All the tests were all two-tailed, and P < 0.05 was thought to be statistically significant.

Clinical characteristics
A total of 410 patients were included in our study.Patients were randomly divided according to the 7:3 ratio, with 288 patients in the training cohort and 122 in the validation cohort.The baseline clinical characteristics of the patients in the training and validation cohorts are listed in Table 1.

Feature selection and radiomics model building
We considered that the ICC value greater than 0.75 indicated a good agreement, then we preformed mRMR and LASSO logistic regression analysis.As a result, a total of We used Mann-Whitney-Wilcoxon U-test to compare the rad-scores between the negative and positive PM in the training and validation cohorts and the difference is statistically significant (P < 0.01) (Fig. 4).

Radiomic nomogram validation
The radiomics signature displayed an AUC for predicting PM status of 0.86 [95% confidence interval (CI), 0.81-0.90]and 0.85 (95% CI, 0.78-0.92) in the training and validation cohorts, respectively (Fig. 5).Logistic regression analysis identified that peritoneal effusion [odds ratio (OR), 17.51], SUVmean (OR, 0.90), and CA125 (OR, 1.004) as independent predictors of the PM (Table 2).After that, the clinical prediction model was constructed and showed an AUC of 0.85 (95% CI, 0.80-0.90)and 0.84 (95% CI, 0.75-0.93).After multivariate logistic regression analysis, the rad-score combined with these clinical factors was used to build a comprehensive predicting model, which we visualized as a radiomics nomogram, and the model showed an AUC of 0.92 (95% CI, 0.89-0.95) in the training cohort and 0.92 (95% CI, 0.86-0.98) in the validation cohort (Fig. 6).The predictive performances of the three models including accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value were shown in Table 3.The calibration curves of our nomogram demonstrated a good consistency between prediction and actual situation (Fig. 7).Comparing to the clinical and radiomics model, the comprehensive model showed a significantly higher discrimination (P < 0.05) in both training and validation cohorts.

Clinical use
The decision curve was produced to compare the net benefits of the radiomics model, the clinical prediction model, and the radiomics nomogram (Fig. 8).The result showed that comprehensive model had a higher overall  net benefit than radiomics and clinical model during the majority range of reasonable threshold probabilities.
PM is the commonest distant metastasis and recurrences location in GC [25], moreover, it is always a clinical difficulty in treatment decision-making, in clinical practice, intraperitoneal chemotherapy is highly recommended for PM patients [26,27], but it is important to treat in a propriate time.Therefore, the construction of preoperative prediction tools is essential.
Several former studies have explored the role of CT  radiomics in PM diagnosis [28][29][30], but the limitation of a single modality is inevitably.Instead, PET/CT allows a higher detection rate of distant metastasis in a variety of tumors [31][32][33].
In the clinical characteristics, peritoneal effusion, CA125, and SUVmean was considered as the independent dictive factors.Peritoneal effusion represents the malfunction of the peritoneum.In GC patients, massive peritoneal effusion indicates more possibility of PM, when visual effusion was found on CT image, diagnostic intra-peritoneal fluid examination could be performed to identify whether there were dissociate cancer cells in abdominal cavity.In our previous study, existence of peritoneal effusion was not collected, after adding this part of data, the clinical predictive model showed a better performance than the former study.(AUC of current study:0.85,0.84 vs. former study 0.76, 0.69 in training and validation cohorts, respectively) Currently, serum tumor markers (CEA, CA199, CA125) are the commonest indexes for malignant tumor screening, several studies have pointed out that these serum markers play an important role in monitoring of GC and predicting the prognosis [34,35].Moreover, Hu et al. [36].found that the tumor markers before and after neoadjuvant chemotherapy have a discriminatory ability for patients with GC, and the normalization of tumor markers was associated with better survival.When CA125 > 17.3 U/ml, there will be higher risk of PM in GC patients [37].Commonly, SUVmax is the most frequently used PET/ CT parameter, it reflects the most active metabolism part about the whole area of tumor.But the SUVmax is not absolutely correlated with the entire tumor burden and its metastasis possibility [38], comparing to our former study, the SUVmean which showed the total metabolism level of the entire tumor was chosen into the prediction model rather than the SUVmax.
Our radiomics signature was consist of 14 radiomics features and obviously correlated with PM.It was an independent predictor of PM (P < 0.01).The selection of features was performed by mRMR and LASSO.The mRMR could reduce feature selection time and improve the classification accuracy significantly, and the LASSO method uses shrinkage procedure, which lead to a more stable feature selection.In our chosen features, GLRLM_ LRLGE, GLZLM_SZE and NGLDM_Contrast in PET and CT were also reported in other literatures.Normally, GLRLM were used to describe the direction, adjacent interval, and variation amplitude in the grayscale [15].Beukinga et al [39].pointed out that the Long-Run Low Gray-level Emphasis (LRLGE) before treatment was superior to SUVmax in predicting the efficacy of neoadjuvant radiotherapy and chemotherapy in advanced esophageal cancer.Intuitively, when GLRLM_LRLGE is wider, the texture of the original image will become coarser.NGLDM reflected the grey-level variability between one voxel and its neighbors in three dimensions, while Contrast measured the intensity difference between neighboring regions.Zhou et al [40].found that radiomic features including NGLDM_Contrast PET combined with genomic factors could predict the prognosis of B-Cell Lymphoma patients receiving Chimeric Antigen Receptor T-cell therapy.Commonly, higher NGLDM_ Contrast was associated with poor prognosis.In our study, both NGLDM_Contrast PET and NGLDM_Contrast CT were chosen, indicating more heterogeneity of GC lesion and higher possibility of PM.Decision curve analysis for three models.
Our nomogram could be easily applied in clinical practice.Table 3 showed the predictive performances of our three models, as we can see that the radiomics model had a higher sensitivity (0.820, 0.842 in training and validation cohorts), while the clinical model held a better specificity (0.949, 0.964 in training and cohorts).By combining the clinical risk factors with the radiomics model, the AUCs of the nomogram were significantly improved in both cohorts (from 0.84-0.86 to 0.92), the performance of current study was also better than our former study (AUC 0.92 vs. 0.90).Besides, the PPV improved a lot in the nomogram (from 0.581 in radiomics training cohort to 0.842), the PPV reflects the possibility of real PM positive among the positive results of our prediction, which meant the combination led to a better accuracy of the PM diagnosis.
The possibility of PM in GC patients will be produced for each patient after the clinical risk factors and radiomics signature acquired from preoperative medical records and medical images.By combining the result of nomogram, a more personalized treatment decision can be made for patients.Our nomogram suggests that a patient with higher CA125 level, higher SUVmean, visual peritoneal effusion and higher radiomics rad-score should accept a diagnostic laparoscopy to identify the PM status, even when the 18 F-FDG PET/CT is negative for PM affectation.The systemic chemotherapy could be used earlier for the PM patients to prolong the survival time and improve their prognosis.
Of course, our study still has several limitations.Firstly, the samples were from a single-center and the models were lack of external validation.Secondly, our study was a retrospective research, thus selective bias cannot be avoided.Thirdly, a part of patients without visible lesions on 18 F-FDG PET images were excluded in our study due to the low uptake of 18 F-FDG and difficulty to set an accurate ROI.Normally, the pathological type of this part of patients would most probably be gastric signet-ring cell carcinoma or the mucinous adenocarcinoma, Chen et al [41].found that the SUVmax of adenocarcinoma is significantly higher than signet-ring cell carcinoma.In clinical practice, such type of GC is highly recommended to gastroscopy to identify the diagnosis due to its possibility of false negative.For these patients, setting ROI on the peritoneum could be a way to predict PM, which is also included in our next work.We did not include Lauren classification in the clinical factors because of the low consistent rate between biopsy and pathology [42].The Lauren classification is consisted of intestinal type, diffuse type, and indetermined type.However, a number of special types of GC have been identified which do not fit into the Lauren types.Therefore, in our future work, we plan to enlarge the size of samples and perform a multi-centers test to validate our results.An 18 F-FDG PET/CT based model whose ROI were set on the peritoneum region will be constructed to predict the PM and the prognosis of the patients.

Conclusion
In conclusion, we constructed a nomogram based on 18 F-FDG PET/CT and clinical risk factors, the nomogram has a good performance to predict PM status of GC patients, and could be a useful tool to assist clinicians in individual clinical decision-making.

( a )
The nomogram of PM prediction in GC patients.Calibration curves of the nomogram in the (b) training cohort and (c) validation cohort.

Table 3
Predictive performances of the models in the training and validation cohorts