Gastric cancer (GC) is the fifth most frequently diagnosed cancer worldwide and the third leading cause of cancer-related deaths. Owing to the large population, approximately 44% of the GC cases worldwide have occurred in China, with an adjusted incidence rate of 20.6/100,000 in 2020. The tumor-lymph node-metastasis (TNM) staging system is the most commonly used benchmark for predicting the long-term survival of patients with GC. However, there are large variations in survival outcomes, even among patients at the same stage and with similar treatment regimens. Moreover, weak correlations between the TNM staging system and recurrence patterns have been reported. These findings suggest that this anatomy-based system provides incomplete prognostic information. Thus, new valuable and robust strategies are warranted to improve outcome prediction and guide personalized treatment and surveillance strategies.
Recently, radiomics, an emerging technology for converting imaging data into a large panel of quantitative features using a large number of automatically applied data-characterization algorithms, has garnered increasing attention. Radiomics enables the noninvasive profiling of tumor heterogeneity by combining high-dimensional mineable features in parallel, which may facilitate a better understanding of tumor behavior. Previous studies have reported the improved performance of CT-based radiomics signature in clinical diagnosis and staging over conventional imaging metrics. More recent studies have also explored the associations of imaging features with the tumor immune microenvironment and response to immunotherapy. In a retrospective study of 1,591 patients with GC, Jiang et al. developed a multiparametric radiomics signature for risk stratification and treatment selection, which may add value to the current staging system. However, few studies have predicted the long-term outcomes of locally advanced GC, and research on radiomics concerning the recurrence patterns is still lacking.
In this study, using the least absolute shrinkage and selection operator (LASSO) Cox regression model, we developed and validated a multiple-feature-based radiomics score to predict the recurrence-free survival (RFS) and overall survival (OS) of patients with locally advanced GC after radical gastrectomy, and assessed its incremental value in the TNM staging system. Furthermore, we explored the potential association between the radiomics score and patterns of recurrence.
MATERIALS AND METHODS
Between January 2009 and June 2015, the data on 513 consecutive patients who underwent radical gastrectomy for GC at two institutions in China [Affiliated People's Hospital of Fujian University of Traditional Chinese Medicine (n = 327) and Qinghai Provincial People's Hospital (n = 186)] were retrospectively reviewed. The inclusion criteria of the patients were as follows: aged 18 to 80 years, histopathologically confirmed primary gastric adenocarcinoma, contrast-enhanced abdominal computed tomography (CT) performed <30 days before surgery, R0 resection, at least 15 lymph nodes harvested, and complete follow-up data. The exclusion criteria of the patients were as follows: M1 or T4b disease, malignant disease of other organs, previous gastrectomy, preoperative anti-tumor treatment, and postoperative death within 3 months. Patients with clinical T1 disease were also excluded from the study because their tumor lesions were difficult to identify using CT. The flow diagram is shown in [https://links.lww.com/SJGA/A75]. Clinicopathological characteristics, including age, sex, American Society of Anesthesiologists (ASA) score, Lauren type, tumor stage, tumor size, lymphovascular invasion, neural invasion, carcinoembryonic antigen (CEA), cancer antigen (CA) 19-9, and adjuvant chemotherapy were collected from the medical records.
All surgical procedures, including standard gastrectomy and D2 lymph node dissection, were performed according to the Japanese Gastric Cancer Treatment Guidelines. The type of reconstruction, including Billroth I gastroduodenostomy, Billroth II gastrojejunostomy, and Roux-en-Y gastrojejunostomy, was determined according to the surgeon's experience and preference. Additionally, delta-shaped anastomosis was utilized for intracorporeal Billroth I reconstruction after laparoscopic distal gastrectomy. The TNM classification (American Joint Committee on Cancer, 8th edition) was used for tumor staging. All patients routinely received 5-fluorouracil (5-FU)-based adjuvant chemotherapy after surgery.
Recurrence was defined as the presence of a biopsy-proven tumor showing adenocarcinoma cells or imaging features that indicate a higher likelihood of tumor recurrence. Recurrences were categorized by the site involved: locoregional, peritoneal, and distant. Although some patients had multiple recurrence episodes, this study analyzed only the initial recurrence episodes, as defined earlier. RFS was defined as the time interval from surgery to recurrence or last follow-up. OS was defined as the time from surgery to death from any cause, or the last follow-up.
All patients were followed up postoperatively every 3 months for 2 years, every 6 months during 3 to 5 years, and annually thereafter, with physical examinations, laboratory tests, and imaging examinations (including chest radiography, abdominal ultrasonography, or abdominopelvic CT). Additionally, an annual endoscopy was recommended. The follow-up period was completed on August 2021. The median follow-up time was 81 months (range: 3–144 months) in the training cohort and 66 months (range: 3–99 months) in the validation cohort.
CT acquisition and feature extraction
Portal venous-phase CT images (thickness: 2.5 mm) were retrieved from the picture archiving and communication system (Carestream, Canada). Details regarding image preprocessing are presented in the Supplementary Materials. For each lesion, a single region of interest was manually delineated on the transverse image section that depicted the largest tumor area using 3D Slicer (version 4.9.0; http://www.slicer.org) by one radiologist with 15 years of experience in abdominal CT interpretation. Feature extraction was performed using the open-source PyRadiomics package (version 2.12; https://pyradiomics.readthedocs.io/en/2.1.2/). Radiomics features were defined in compliance with the Imaging Biomarker Standardization Initiative. We extracted 851 radiomic features [https://links.lww.com/SJGA/A80].
Feature selection and model construction
We performed a three-step procedure to select robust prognostic factors from the candidate radiomic features in the training cohort. First, we applied univariable Cox regression analysis to identify significant predictors of RFS with a P value less than 0.05 [https://links.lww.com/SJGA/A80]. A total of 260 features were screened out for the next step. Second, we calculated the Pearson linear correlation coefficient for each of the two features [https://links.lww.com/SJGA/A81]. For those feature pairs with a correlation coefficient over 0.90, we only kept the one with a lower P value. Thus, 153 features were excluded. Finally, the LASSO Cox regression model was utilized to identify the optimal combination of radiomic features via the 1-SE (standard error) criteria [https://links.lww.com/SJGA/A82]. Five features were determined in the last step, and their associations with RFS were investigated using restricted cubic splines [https://links.lww.com/SJGA/A83]. The radiomics score was then generated via a linear combination of selected features weighted by their respective coefficients. Additional details are presented in the Supplementary Materials.
Validation of the prediction model
The prediction model was validated by measuring both discrimination and calibration. Harrell's concordance index (C-index) was measured to quantify discrimination performance. Calibration curves were generated to compare the predicted survival with actual survival. The Akaike information criterion (AIC) within the Cox regression model was used to compare performance between the two models; smaller AIC values represent more optimistic prognostic stratification. The likelihood ratio of the Chi-square score was calculated using Cox regression to measure homogeneity; a higher likelihood ratio of the Chi-square score indicates better homogeneity. A net reclassification improvement (NRI) calculation was used to quantify the improvement in usefulness. We also performed a decision curve analysis to determine the clinical usefulness of the model by quantifying the net benefits at different threshold probabilities.
Continuous variables are presented as mean ± standard deviation (SD) values, or median ± interquartile range (IQR) values, and categorical variables are presented as frequencies and percentages. The differences between the groups were assessed using the t-test, one-way analysis of variance (ANOVA), the Mann–Whitney test, or the Chi-squared test, as appropriate. A minimum P value approach was used to evaluate the optimal cut-off values of the radiomics score using the X-tile program (3.6.1 software 20, http://medicine.yale.edu/lab/rimm/research/software.aspx). Survival analysis was performed using the Kaplan–Meier survival curves, and a log-rank test was used to determine significance. Univariate and multivariate analyses were performed using the Cox proportional hazards model and a logistic regression model. Variables with P < 0.05 in the univariate analysis were included in the subsequent multivariate analysis. A nomogram consisting of the radiomics score and clinicopathological factors was created to translate model parameter estimates into a visual scoring system to calculate the estimated survival probability.
Statistical analyses were performed using SPSS version 22.0 for Windows (SPSS Inc., Chicago, IL, USA) and the R software (version 4.1.1; R Foundation for Statistical Computing; https://www.r-project.org/). All tests were two-sided, with a significance level of P < 0.05.
The clinicopathological characteristics of the training and validation cohorts are presented in Table 1. There were 244 (74.6%) men and 83 (25.4%) women in the training cohort (n = 327). The mean age at diagnosis was 60.6 years (SD 11.5 years). Of the 186 patients in the validation cohort, 133 patients (71.5%) were men and the mean age was 58.8 years (SD 11.6 years). Patients in the training cohort were more likely to undergo laparoscopic gastrectomy (74.6% vs. 52.2%, P < 0.001) and Billroth II gastrojejunostomy (56.6% vs. 31.7%, P < 0.001) than those in the validation cohort. No significant differences were found in the other characteristics between the two cohorts (P > 0.05). In the training cohort, the 5-year RFS and OS rates were 61.1% (95% CI 56.0–66.7%) and 62.1% (95% CI 57.0–67.6%), respectively, and in the validation cohort, the 5-year RFS and OS rates were 56.6% (95% CI 49.8–64.3%) and 56.8% (95% CI 50.1–64.4%), respectively.
Univariate and multivariate analyses
Using the LASSO model, a prognostic score was constructed based on five radiomics features that were selected from the 105 potential predictors in the training cohort [Figure S3]. In the univariate analysis, age (P = 0.006 and 0.001, respectively), pT stage (P both <0.001), pN stage (P both <0.001), tumor size (P both <0.001), CEA (P = 0.049 and 0.011, respectively), CA19-9 (P both <0.001), and radiomics score (P both <0.001) were significantly associated with RFS and OS. In the multivariate analysis, the radiomics score remained an independent prognostic factor for both RFS (HR 2.978, 95% CI 1.914–4.634, P < 0.001) and OS (HR 2.684, 95% CI 1.729–4.168, P < 0.001). In the validation cohort, a higher radiomics score was independently associated with worse RFS (HR 2.990, 95% CI 1.441-6.203, P = 0.003) and OS (HR 3.029, 95% CI 1.426-6.434, P = 0.004, Table 2, https://links.lww.com/SJGA/A84).
Prognostic value of the radiomics score
In the training cohort, the C-indexes of the radiomics score for RFS and OS were 0.740 (95% CI 0.702–0.778) and 0.726 (95% CI 0.688–0.766), respectively; in the validation cohort, the C-indexes for RFS and OS were 0.734 (95% CI 0.685–0.784) and 0.730 (95% CI 0.679–0.782), respectively. The optimum cutoffs for the radiomics score generated by the X-tile plot were − 0.16 and 0.32 [https://links.lww.com/SJGA/A85]. Patients were then classified into a low-radiomics score group (low-RS, radiomics score <−0.16), a medium-radiomics score group (medium-RS, −0.16 ≤ radiomics score < 0.32), and a high-radiomics score group (high-RS, radiomics score ≥ 0.32). More advanced disease (P < 0.001), larger tumors (P < 0.001), higher CEA levels (P < 0.001), and higher CA19-9 levels (P = 0.045) were significantly related to high RS status [https://links.lww.com/SJGA/A86]. In the training cohort, the 5-year RFS and OS rates were 88.1% (95% CI 82.3–94.4%) and 89.2% (95% CI 83.6–95.2%), respectively, for the low-RS group; 53.4% (95% CI 46.3–61.6%) and 51.2% (95% CI 44.2–59.3%), respectively, for the medium-RS group; 17.4% (95% CI 9.3–32.6%) and 23.9% (95% CI 14.3–40.0%), respectively, for the high-RS group (P < 0.001 for all, Figure 1a). In the validation cohort, the radiomics score also classified the patients into three risk groups with significantly different RFS and OS (P < 0.001, [Figure 1b]). We further performed subgroup analyses for the training and validation cohorts. The radiomics score remained a statistically significant prognostic predictor in different patient populations (P all <0.001, https://links.lww.com/SJGA/A87 and https://links.lww.com/SJGA/A76).
Construction of the Nomogram
To provide a quantitative method to predict the probability of the 1-, 3-, and 5-year RFS, a nomogram that integrated the radiomics score and clinicopathological factors was established [Figure 2a]. Calibration curves showed that the nomogram performed well compared with actual observations in the training and validation cohorts [Figure 2b and c]. Decision curve analysis showed that the nomogram had better clinical utility than the TNM staging system across the majority of the range of reasonable threshold probabilities [Figure 3]. Compared with the TNM staging system alone, the nomogram showed significantly better discrimination performance with higher C-indexes (95% CIs) in the training cohort (RFS: 0.791 [0.759–0.824] vs. 0.762 [0.725–0.800]; OS: 0.782 [0.748–0.816] vs. 0.747 [0.710–0.784]; P < 0.05). The nomogram also had a lower AIC (RFS: 1345.4 vs. 1369.8; OS: 1454.3 vs. 1482.3) and a higher likelihood ratio of the Chi-square (RFS: 145.8 vs. 121.4%; OS: 133.3 vs. 105.3%) than the TNM staging system, which represented better goodness of fit and better predictive homogeneity. Moreover, the classification accuracy for survival outcomes of the nomogram was improved compared with the TNM staging system alone, and the NRIs for RFS and OS were 30.1% (95% CI 4.6–66.8%) and 38.4% (95% CI 18.6–64.0%), respectively. Similar results were obtained for the validation cohort [https://links.lww.com/SJGA/A77].
Association between the Radiomics Score and recurrence patterns
A total of 217 patients (42.3%) experienced disease recurrence in the combined training and validation cohorts. Of these, 186 (85.7%) patients experienced recurrence involving a single area, 30 (9.7%) experienced recurrence involving two areas, and one (0.5%) experienced recurrence involving all three areas. Distant, peritoneal, and locoregional recurrences occurred in 124 (57.1%), 80 (36.9%), and 45 (20.7%) patients, respectively [Figure 4a]. When stratified by the radiomics score, the proportion of locoregional recurrence was highest in the low-RS group (60.9%), the proportion of peritoneal recurrence was highest in the medium-RS group (47.1%), and the proportion of distant recurrence was highest in the high RS group (75.9%, Figure 4b). To confirm the associations between the radiomics score and recurrence patterns in different populations, we performed a subgroup analysis and found similar results [https://links.lww.com/SJGA/A78a-f. Multivariate analysis also revealed that the radiomics score was an independent predictor of recurrence patterns [https://links.lww.com/SJGA/A79].
In this study, we built a five-texture features-based radiomics score that was significantly associated with RFS and OS and was an independent prognostic factor in patients with locally advanced GC, who underwent radical gastrectomy. Moreover, the radiomic score could categorize patients into low-, medium-, and high-RS groups with significant differences in the patterns of recurrence. When the patients were stratified by clinicopathological factors, the radiomics score continued to provide significant predictive power for survival outcomes and recurrence patterns. Incorporating the radiomics score into the radiomics-based nomogram provided better predictive power and improved classification accuracy in the estimation of RFS and OS than the TNM staging system alone. These results suggest that this multiple-feature-based score is a reliable, noninvasive, and low-cost tool for preoperative individual prediction of long-term outcomes.
Intratumor heterogeneity has been reported to correlate with worse outcomes and manifests on multiple spatial scales. Radiomics, which extends the analysis of individual quantitative radiomics features to an "omics"-based approach, allows us to assess intra-tumor heterogeneity quantitatively on a macroscopic tissue scale and has become a hot topic in recent years. Previous studies have supported that phenotypic and proteomic information of tumors can be deduced from radiomics features. For example, in a retrospective study of 1591 patients with GC, Jiang et al. developed a radiomics signature that consisted of 19 selected features for the prediction of long-term survival and chemotherapeutic benefits. Moreover, Jiang et al. demonstrated that the radiomics signature was a reliable tool for the evaluation of the immunoscore and retained its prognostic significance. In the present study, the radiomics score independently and effectively predicted RFS and OS and stratified the patients, by separating them into three risk groups with significantly different outcomes, which also supports the hypothesis that the radiomics signature has the potential to capture intratumoral heterogeneity in a noninvasive method. For patients with a high radiomics score, more effective systemic therapies, such as perioperative chemotherapy, should be comprehensively considered.
The TNM staging system maintained by the American Joint Committee on Cancer (AJCC) is the most widely used criterion for predicting survival in patients with GC. This system stratified M0 GC into seven risk groups according to the depth of invasion and the number of lymph node metastases. However, other factors, such as genetic, cellular, and clinicopathological characteristics, should be considered when predicting individualized survival according to the previous studies. As shown in the present study, the radiomics score successfully identified high-risk patients with poor survival outcomes in stage Ib and identified low-risk patients with good survival outcomes in stage III. These findings suggested that scoring could reinforce the prognostic value of TNM staging. A nomogram combining the radiomics score, T stage, N stage, and CA19-9 level was then constructed with a high C-index and positive NRI. This model enabled a more accurate prediction of individual outcomes and better clinical utility than the traditional staging system, which would be beneficial for scheduling personalized treatment and follow-up strategies.
Data on the recurrence patterns provide valuable information for conducting appropriate therapies and effective follow-up examinations. Clinicopathologic factors associated with recurrence patterns have been extensively investigated. In 386 patients with GC who developed recurrence after resection, patterns of recurrence varied significantly based on the Lauren histologic type. In a larger study involving 656 patients with recurrent GC, the pattern of initial recurrence was significantly different according to the pathologic stage. However, there is still no consensus on the optimal assessment tools for recurrence patterns. In this study, a low-, medium-, and high-radiomics score was significantly associated with an increased risk of locoregional, peritoneal, and distant recurrence. Subgroup and multivariate analyses also confirmed these findings. To the best of our knowledge, this is the first study to demonstrate the association between the radiomics signature and patterns of recurrence, which can be conveniently used to facilitate postoperative surveillance.
Our study has several limitations. First, as a retrospective study, it may have been subject to selection bias. Second, because of the relatively small number of patients in the training cohort, the developed radiomics score may not be the best and most effective. Finally, the relationships between imaging signatures and genomic sequencing were not assessed. Nevertheless, this scoring method may provide insights that can improve personalized medicine and may serve as a potential tool to guide individual care in patients with GC.
In conclusion, the radiomics score is a reliable tool that can effectively predict the survival outcomes of patients with GC and improve the prognostic value of the TNM staging system. Moreover, the radiomics score may be a useful predictor of recurrence patterns. A large-scale, prospective, well-designed study is warranted to evaluate the clinical feasibility of these results.
Financial support and sponsorship
This study was funded by the Social Development Guiding (Key) Project of Fujian Provincial Department of Science and Technology (No. 2019Y0034) and the Key project of National Traditional Chinese Medicine Research Base (No.JDZX201903).
Conflicts of interest
There are no conflicts of interest.
- Supplementary Methods
- Supplementary Reference
- Supplementary Figures
- Supplementary Tables
CT image acquisition
All patients underwent contrast-enhanced abdominal CT using the multidetector row CT (MDCT) systems (Discovery CT750 HD scanner, GE Healthcare, USA). The acquisition parameters were as follows: 120 kV; 100-120 mAs; 0.5-second rotation time; detector collimation: 64×0.625 mm; field of view, 350×350 mm; matrix, 512×512. All patients drank 600–1000 ml of warm water to distend the stomach prior to CT examination. After routine non-enhanced CT, arterial and portal venous-phase contrast-enhanced CT were performed after delays of 28 s and 60 s, following intravenous administration of iodinated contrast material (Ultravist 370, Bayer Schering Pharma, Berlin, Germany), at a rate of 3.0 or 3.5 ml/s with a pump injector (Ulrich CT Plus 150, Ulrich Medical, Ulm, Germany). Contrast-enhanced CT was reconstructed with a reconstruction thickness of 2.5 mm. Portal venous phase CT images were retrieved from the picture archiving and communication system (PACS) (Carestream, Canada) for image feature extraction because of well differentiation of the tumor tissue from the adjacent normal bowel wall.
The target lesions were manually segmented in 3D using the ITK-SNAP software by one radiologist with 11 years of clinical experience in abdominal CT interpretation. The radiologist was blind to the clinicopathological data but was aware that the patients had gastric cancer. Radiomics feature extraction was performed using the open-source platform Pyradiomics (version 2.2.0, https://pyradiomics.readthedocs.io/), which enables quantitative features to be extracted from HRCT images. Extracted from Pyradiomics were 14 shape features, 18 first-order features, 22 gray-level co-occurrence matrix (GLCM) features, 16 gray-level size zone matrix (GLSZM) features, 16 gray-level run length matrix (GLRLM) features, and 14 gray-level dependence matrix (GLDM) features. Radiomic features were calculated based on the original image and after wavelet filtering, yielding eight additional image types based on the application of wavelet-based high-pass or low-pass filters to each of the three dimensions. In total, 851 radiomics features [14 shape features + 93 other features × 9 image types] were generated for each nodule. A detailed list of the extracted features is provided in [https://links.lww.com/SJGA/A75].
The values of the extracted features from the training cohort were standardized with z scores; the feature values of the validation cohort were then standardized to z scores by using the mean and standard deviation values derived from the training cohort.
Model Construction using LASSO Cox Regression Model
The least absolute shrinkage and selection operator method (LASSO) is a popular method for regression of high-dimensional predictors. The method uses an L1 penalty to shrink some regression coefficients to exactly zero. LASSO has been extended and broadly applied to the Cox proportional hazard regression model for survival analysis with high-dimensional data. We selected λ via 1-SE (standard error) criteria, i.e., the optimal λ is the largest value for which the partial likelihood deviance is within one SE of the smallest value of partial likelihood deviance. We also plotted the Harrell's concordance index (C-index) versus log (λ), where λ is the tuning parameter. A value of λ= 0.002365175 with log (λ) = -2.62613672 was selected by cross-validation via the 1-SE criteria. A vertical line was drawn at log (λ) = -2.567236, which corresponded to the optimal value λ= 0.07674741 (Figure S3). The optimal tuning parameter resulted in five non-zero coefficients. The following nineteen features were selected in the LASSO Cox regression model: originalshapeMaximum2DDiameterSlice, wavelet-HHHglcmImc2, wavelet-HHLfirstorderTotalEnergy, wavelet-LLLgldmGrayLevelNonUniformity, and originalfirstorder10Percentile, with coefficients 0.111417, 0.050083, 0.073731, 0.195280, and 0.011622.
- Yushkevich PA, Piven J, Hazlett HC, et al. User-guided 3D active contour segmentation of anatomical structures: Significantly improved efficiency and reliability. Neuroimage. 2006; 31: 1116-1128.
- van Griethuysen JJM, Fedorov A, Parmar C, et al. Computational Radiomics System to Decode the Radiographic Phenotype. Cancer Res. 2017; 77: e104-e107.
- Tibshirani R. The lasso method for variable selection in the Cox model. Stat Med. 1997; 16: 385-395.
- Tibshirani R. Regression shrinkage and selection via the lasso: a retrospective. Journal of the Royal Statistical Society Series B-Statistical Methodology. 2011; 73: 273-282.
1. Bray F, Ferlay J, Soerjomataram I, Siegel RL, Torre LA, Jemal A. Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries CA Cancer J Clin. 2018;68:394–424
2. Yang L, Ying X, Liu S, Lyu G, Xu Z, Zhang X, et al Gastric cancer: Epidemiology, risk factors and prevention strategies Chin J Cancer Res. 2020;32:695–704
3. Han DS, Suh YS, Kong SH, Lee HJ, Choi Y, Aikou S, et al Nomogram predicting long-term survival after d2 gastrectomy for gastric cancer J Clin Oncol. 2012;30:3834–40
4. Hirabayashi S, Kosugi S, Isobe Y, Nashimoto A, Oda I, Hayashi K, et al Development and external validation of a nomogram for overall survival after curative resection in serosa-negative, locally advanced gastric cancer Ann Oncol. 2014;25:1179–84
5. Jiang Y, Li T, Liang X, Hu Y, Huang L, Liao Z, et al Association of adjuvant chemotherapy with survival in patients with stage ii or iii gastric cancer JAMA Surg. 2017;152:e171087
6. D'Angelica M, Gonen M, Brennan MF, Turnbull AD, Bains M, Karpeh MS. Patterns of initial recurrence in completely resected gastric adenocarcinoma Ann Surg. 2004;240:808–16
7. Lee JH, Chang KK, Yoon C, Tang LH, Strong VE, Yoon SS. Lauren histologic type is the most important factor associated with pattern of recurrence following resection of gastric adenocarcinoma Ann Surg. 2018;267:105–13
8. Lambin P, Rios-Velazquez E, Leijenaar R, Carvalho S, van Stiphout RG, Granton P, et al Radiomics: Extracting more information from medical images using advanced feature analysis Eur J Cancer. 2012;48:441–6
9. Aerts HJ, Velazquez ER, Leijenaar RT, Parmar C, Grossmann P, Carvalho S, et al Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach Nat Commun. 2014;5:4006 Erratum in: Nat Commun. 2014 5 4644
10. Gillies RJ, Kinahan PE, Hricak H. Radiomics: Images are more than pictures, they are data Radiology. 2016;278:563–77
11. Bi WL, Hosny A, Schabath MB, Giger ML, Birkbak NJ, Mehrtash A, et al Artificial intelligence in cancer imaging: Clinical challenges and applications CA Cancer J Clin. 2019;69:127–57
12. Ma Z, Fang M, Huang Y, He L, Chen X, Liang C, et al CT-based radiomics signature for differentiating Borrmann type IV gastric cancer from primary gastric lymphoma Eur J Radiol. 2017;91:142–7
13. Li J, Fang M, Wang R, Dong D, Tian J, Liang P, et al Diagnostic accuracy of dual-energy CT-based nomograms to predict lymph node metastasis in gastric cancer Eur Radiol. 2018;28:5241–9
14. Wang Y, Liu W, Yu Y, Liu JJ, Jiang L, Xue HD, et al Prediction of the depth of tumor invasion in gastric cancer: Potential role of CT radiomics Acad Radiol. 2020;27:1077–84
15. Jiang Y, Wang H, Wu J, Chen C, Yuan Q, Huang W, et al Noninvasive imaging evaluation of tumor immune microenvironment to predict outcomes in gastric cancer Ann Oncol. 2020;31:760–8
16. Xu Y, Su GH, Ma D, Xiao Y, Shao ZM, Jiang YZ. Technological advances in cancer immunity: From immunogenomics to single-cell analysis and artificial intelligence Signal Transduct Target Ther. 2021;6:312
17. Huang W, Jiang Y, Xiong W, Sun Z, Chen C, Yuan Q, et al Noninvasive imaging of the tumor immune microenvironment correlates with response to immunotherapy in gastric cancer Nat Commun. 2022;13:5095
18. Jiang Y, Chen C, Xie J, Wang W, Zha X, Lv W, et al Radiomics signature of computed tomography imaging for prediction of survival and chemotherapeutic benefits in gastric cancer EbioMedicine. 2018;36:171–82
19. Japanese Gastric Cancer Association. . Japanese gastric cancer treatment guidelines 2010 (ver. 3) Gastric Cancer. 2011;14:113–23
20. Japanese Gastric Cancer Association. . Japanese gastric cancer treatment guidelines 2014 (ver. 4) Gastric Cancer. 2017;20:1–19
21. Kanaya S, Gomi T, Momoi H, Tamaki N, Isobe H, Katayama T, et al Delta-shaped anastomosis in totally laparoscopic Billroth I gastrectomy: New technique of intraabdominal gastroduodenostomy J Am Coll Surg. 2002;195:284–7
22. Amin MB, Edge S, Greene F, Byrd DR, Brookland RK, Washigton MK, et al AJCC cancer staging manual. 2016eighth ed New York(NY) Springer
23. Gonen M, Heller G. Concordance probability and discriminatory power in proportional hazards regression Biometrika. 2005;92:965–70
24. Awad AM. Properties of the Akaike information criterion Microelectron Reliab. 1996;36:457–64
25. Yoon HM, Ryu KW, Nam BH, Cho SJ, Park SR, Lee JY, et al Is the new seventh AJCC/UICC staging system appropriate for patients with gastric cancer? J Am Coll Surg. 2012;214:88–96
26. Pencina MJ, D'Agostino RB Sr, Steyerberg EW. Extensions of net reclassification improvement calculations to measure usefulness of new biomarkers Stat Med. 2011;30:11–21
27. Vickers AJ, Cronin AM, Elkin EB, Gonen M. Extensions to decision curve analysis, a novel method for evaluating diagnostic tests, prediction models and molecular markers BMC Med Inform Decis Mak. 2008;8:53
28. Camp RL, Dolled-Filhart M, Rimm DL. X-tile: A new bio-informatics tool for biomarker assessment and outcome-based cut-point optimization Clin Cancer Res. 2004;10:7252–9
29. Razzak M. Genetics: New molecular classification of gastric adenocarcinoma proposed by The Cancer Genome Atlas Nat Rev Clin Oncol. 2014;11:499
30. McLean MH, El-Omar EM. Genetics of gastric cancer Nat Rev Gastroenterol Hepatol. 2014;11:664–74
31. Jiang Y, Zhang Q, Hu Y, Li T, Yu J, Zhao L, et al Immunoscore signature: A prognostic and predictive tool in gastric cancer Ann Surg. 2018;267:504–13
32. Seo N, Han K, Hyung WJ, Chung YE, Park CH, Kim JH, et al Stratification of postsurgical computed tomography surveillance based on the extragastric recurrence of early gastric cancer Ann Surg. 2020;272:319–25
33. Tang Y, Lin J, Lin J, Wang J, Lu J, Chen Q, et al Reappraise role of lymph node status in patterns of recurrence following curative resection of gastric adenocarcinoma Chin J Cancer Res. 2021;33:331–42
34. Kim JH, Lee HH, Seo HS, Jung YJ, Park CH. Stage-specific difference in timing and pattern of initial recurrence after curative surgery for gastric cancer Surg Oncol. 2019;30:81–6