Advertisement

CT-Based Radiomic Score: A Risk Stratifier in Far-Advanced Gastric Cancer Patients

Open AccessPublished:January 06, 2023DOI:https://doi.org/10.1016/j.acra.2022.12.034

      Objectives

      To prolong the survival, the value of a computed tomography-based radiomic score (RS) in stratifying survival and guiding personalized chemotherapy strategies in far-advanced gastric cancer (FGC) was investigated.

      Materials and Methods

      This retrospective multicenter study enrolled 283 FGC patients (cT4a/bNxM0-1) from three centers. Patients from one center were randomly divided into the training (n = 166) and internal validation (n = 83) cohorts, whereas the external validation cohort (n = 34) consisted of patients from the two other centers. The RS was calculated for each patient to predict progression-free survival (PFS). Features from the primary tumor and main metastasis (peritoneum, liver, and lymph node) were integrated in the training cohort and then validated for its ability to stratify PFS and overall survival (OS) in the validation cohort. The association between the RS and efficacy of neoadjuvant intraperitoneal and systemic (NIPS) therapy was also explored.

      Results

      The RS demonstrated a favorable prognostic ability to predict PFS in all cohorts (training: C-index 0.83, 95% confidence interval [CI]: 0.788-0.872; internal validation: C-index 0.75, 95% CI: 0.682-0.818; external validation: C-index 0.76, 95% CI: 0.669-0.851; all p < 0.05), as well as an excellent ability to stratify the PFS and OS in both the whole population and metastatic subgroups (p < 0.05). Patients with a low score were more likely to undergo surgery after perioperative chemotherapy (p < 0.05). Furthermore, only high-scoring patients with peritoneal metastasis benefited from NIPS.

      Conclusion

      The RS may be an effective risk stratifier for the outcomes of FGC patients and may be used to select patients who can benefit from NIPS therapy.

      Key Words

      Abbreviations:

      AFP (alpha-fetoprotein), AUC (area under the curve), C-index (concordance index), CA199 (carbohydrate antigen 199), CA724 (carbohydrate antigen 724), CA125 (carbohydrate antigen 125), CEA (carcinoembryonic antigen), CI (confidence interval), cM (clinical distant metastasis), CT (computed tomography), DCA (decision curve analysis), FGC (far-advanced gastric cancer), HR (hazard ratios), ICC (intraclass coefficient), KM (Kaplan–Meier), LM (liver metastasis), LNM (lymph node metastasis), NIPS (neoadjuvant intraperitoneal and systemic chemotherapy), OS (overall survival), PFS (progression-free survival), PM (peritoneal metastasis), PS score (performance status score), ROC (receiver operating characteristic), ROI (regions of interest), RS (radiomic score), TNM (tumor node-metastasis), UICC (Union for International Cancer Control)

      INTRODUCTION

      Gastric cancer ranks fifth among all cancers worldwide in terms of incidence (approximately 5.7%) and third in mortality (approximately 8.2%) (
      • Siegel RL
      • Miller KD
      • Fuchs HE
      • et al.
      Cancer statistics, 2021.
      ). About two-thirds of patients have advanced gastric cancer when first diagnosed clinically, 30%–40% of whom have unresectable far-advanced gastric cancer (FGC) with a poor prognosis (
      • Cheng XJ
      • Lin JC
      • Tu SP.
      Etiology and prevention of gastric cancer.
      ). For patients with unresectable FGC who respond well to chemotherapy, radical resection is recommended, but the success rate is approximately 22%–35% (
      • Saito M
      • Kiyozaki H
      • Takata O
      • et al.
      Treatment of stage IV gastric cancer with induction chemotherapy using S-1 and cisplatin followed by curative resection in selected patients.
      ,
      • Uemura N
      • Kikuchi S
      • Sato Y
      • et al.
      A phase II study of modified docetaxel, cisplatin, and S-1 (mDCS) chemotherapy for unresectable advanced gastric cancer.
      ). However, the overall disease progression and recurrence rates after resection are still around 25%–50%, which in turn indicates lower survival (
      • Bilici A
      • Selcukbiricik F.
      Prognostic significance of the recurrence pattern and risk factors for recurrence in patients with proximal gastric cancer who underwent curative gastrectomy.
      ,
      • Nakagawa M
      • Kojima K
      • Inokuchi M
      • et al.
      Patterns, timing and risk factors of recurrence of gastric cancer after laparoscopic gastrectomy: reliable results following long-term follow-up.
      ). Therefore, identifying effective strategies for prolonging survival in FGC remains a challenge to clinicians. An efficient model or scoring tool for predicting survival in unresectable FGC would be crucial to improve individualized treatments.
      Peritoneal metastasis (PM) is the most frequent type of gastric cancer recurrence and metastasis, accounting for approximately 60% of cancer-related deaths (
      • Macri A
      • Morabito F.
      The use of intraperitoneal chemotherapy for gastric malignancies.
      ). The median survival in FGC with PM is only 3.1–10.6 months, whereas the 1-year survival is 16.0%–40.7% (
      • Fujitani K
      • Yang HK
      • Kurokawa Y
      • et al.
      Randomized controlled trial comparing gastrectomy plus chemotherapy with chemotherapy alone in advanced gastric cancer with a single non-curable factor: Japan Clinical Oncology Group Study JCOG 0705 and Korea Gastric Cancer Association Study KGCA01.
      ,
      • Yamaguchi H
      • Kitayama J
      • Ishigami H
      • et al.
      A phase 2 trial of intravenous and intraperitoneal paclitaxel combined with S-1 for treatment of gastric cancer with macroscopic peritoneal metastasis.
      ). Evidence from clinical trials indicate the effectiveness of neoadjuvant intraperitoneal and systemic chemotherapy (NIPS; oral administration of S-1 + intravenous and intraperitoneal chemotherapy with paclitaxel), which has drawn extensive attention in the treatment of FGC with PM (
      • Ishigami H
      • Kitayama J
      • Kaisaki S
      • et al.
      Phase II study of weekly intravenous and intraperitoneal paclitaxel combined with S-1 for advanced gastric cancer with peritoneal metastasis.
      ,
      • Gong Y
      • Wang P
      • Zhu Z
      • et al.
      Benefits of Surgery After Neoadjuvant Intraperitoneal and Systemic Chemotherapy for Gastric Cancer Patients With Peritoneal Metastasis: A Meta-Analysis.
      ). However, not all patients with PM could benefit from NIPS. Particularly, some patients even suffer from serious treatment-related adverse reactions, such as diarrhea, fever, leukopenia, neutropenia, anemia, and thrombocytopenia, (
      • Ishigami H
      • Fujiwara Y
      • Fukushima R
      • et al.
      Phase III Trial Comparing Intraperitoneal and Intravenous Paclitaxel Plus S-1 Versus Cisplatin Plus S-1 in Patients With Gastric Cancer With Peritoneal Metastasis: PHOENIX-GC Trial.
      ). In addition, the effects of NIPS therapy on patients without peritoneal metastasis is unknown. Therefore, tools that help identify patients with FGC who could benefit from NIPS therapy are fundamental.
      Computed tomography (CT) has been widely used in diagnosing and staging gastric cancer. It can also provide imaging biomarkers for predicting prognosis (
      • Wang S
      • Feng C
      • Dong D
      • et al.
      Preoperative computed tomography-guided disease-free survival prediction in gastric cancer: a multicenter radiomics study.
      ). However, CT images merely provide structural information and cannot directly provide information on prognosis-related pathophysiological features (
      • Nakamoto T
      • Haga A
      • Takahashi W.
      [An Introduction to Radiomics: Toward a New Era of Precision Medicine].
      ). Radiomics has emerged as a noninvasive comprehensive analytical method for describing tumor phenotypes and pathophysiology using minable features extracted from numerous medical images (
      • Nakamoto T
      • Haga A
      • Takahashi W.
      [An Introduction to Radiomics: Toward a New Era of Precision Medicine].
      ), thereby significantly extending the value of data in CT images. As reported, CT radiomics has significantly improved the diagnostic accuracy, response to neoadjuvant chemotherapy, and prognosis prediction (
      • Zhang L
      • Dong D
      • Zhang W
      • et al.
      A deep learning risk prediction model for overall survival in patients with gastric cancer: A multicenter study.
      ,
      • Sun KY
      • Hu HT
      • Chen SL
      • et al.
      CT-based radiomics scores predict response to neoadjuvant chemotherapy and survival in patients with gastric cancer.
      ,
      • Xie W
      • Jiang Z
      • Zhou X
      • et al.
      Quantitative Radiological Features and Deep Learning for the Non-Invasive Evaluation of Programmed Death Ligand 1 Expression Levels in Gastric Cancer Patients: A Digital Bopsy Study.
      ,
      • Liu S
      • Qiao X
      • Xu M
      • et al.
      Development and validation of multivariate models integrating preoperative clinicopathological parameters and radiographic findings based on late arterial phase CT images for predicting lymph node metastasis in gastric cancer.
      ). Nevertheless, studies on radiomic analysis of gastric cancer have mostly focused on patients with either early-stage or locally resectable advanced gastric cancer following neoadjuvant chemotherapy; few have investigated the value of radiomic analysis in identifying predictors of survival in unresectable FGC (
      • Liao F
      • Guo X
      • Lu X
      • Dong W.
      A validated survival nomogram for early-onset diffuse gastric cancer.
      ,
      • Shin J
      • Lim JS
      • Huh YM
      • et al.
      A radiomics-based model for predicting prognosis of locally advanced gastric cancer in the preoperative setting.
      ,
      • Li W
      • Zhang L
      • Tian C
      • et al.
      Prognostic value of computed tomography radiomics features in patients with gastric cancer following curative resection.
      ). Furthermore, previous studies have only focused on the features of primary tumors without metastasis, which possibly limits the accuracy of estimation, especially in FGC. An imaging prediction tool could achieve improved prediction of patient survival by incorporating metastasis data, including PM, liver metastasis (LM), and lymph node metastasis (LNM), which have prognostic and predictive value (
      • Wei XL
      • Xu JY
      • Wang DS
      • et al.
      Baseline lesion number as an efficacy predictive and independent prognostic factor and its joint utility with TMB for PD-1 inhibitor treatment in advanced gastric cancer.
      ,
      • Granieri S
      • Altomare M
      • Bruno F
      • et al.
      Surgical treatment of gastric cancer liver metastases: Systematic review and meta-analysis of long-term outcomes and prognostic factors.
      ,
      • Ter Veer E
      • van Kleef JJ
      • Schokker S
      • et al.
      Prognostic and predictive factors for overall survival in metastatic oesophagogastric cancer: A systematic review and meta-analysis.
      ).
      In this study, we investigated the association of imaging features of both the primary tumor and main metastatic lesions with the prognosis of patients with FGC. We developed and validated radiomic score (RS) as predictors of progression-free survival (PFS) and overall survival (OS). We further evaluated the value of radiomic score in stratifying survival risk and its ability to identify patients who can benefit from NIPS therapy.

      MATERIALS AND METHODS

      Patients

      This retrospective multicenter study was conducted in accordance with the Declaration of Helsinki, and the study protocol was approved by the ethics committee of each of the three participating hospitals. The requirement for informed consent was waived.
      A total of 283 consecutive patients with FGC (cT4a/bNxM0-1) were enrolled from three hospitals in Shanghai, China. Data on the patients from January 2017 to May 2020 were collected. Supplemental Material 1 summarizes the details of the pretreatment assessment. FGC was diagnosed by biopsy specimen analysis, upper gastrointestinal endoscopy, CT, positron emission tomography, and magnetic resonance imaging. Chemotherapy was administered following treatment regimens are recommended in the Chinese Society of Clinical Oncology guidelines for gastric cancer treatment (
      • Wang FH
      • Zhang XT
      • Li YF
      • et al.
      The Chinese Society of Clinical Oncology (CSCO): clinical guidelines for the diagnosis and treatment of gastric cancer, 2021.
      ). We randomly divided patients from Center One into the training (n = 166) and internal validation (n = 83) cohorts, whereas the patients from Centers Two and Three comprised the external validation cohort (n = 34). Detailed descriptions of the inclusion and exclusion criteria, as well as of the recruitment process, are shown in Figure 1. Detailed information on the inclusion and exclusion criteria is shown in Supplementary Material 2.
      Figure 1
      Figure 1Recruitment pathway for far-advanced gastric cancer patients for the training cohort and validation cohorts. CT, computed tomography.
      Clinical data from medical records were collected, including sex; age; performance status score (PS score); primary tumor location; clinical stage; and tumor serological indicators, such as alpha-fetoprotein (AFP), carcinoembryonic antigen (CEA), carbohydrate antigen 199 (CA199), carbohydrate antigen 724 (CA724), and carbohydrate antigen 125 (CA125). For the clinical staging of gastric cancer, we applied the eighth edition of the tumor–node–metastasis (TNM) classification published by the Union for International Cancer Control (UICC) (
      • Sano T
      • Coit DG
      • Kim HH
      • et al.
      Proposal of a new stage grouping of gastric cancer for TNM classification: International Gastric Cancer Association staging project.
      ). Information on the evaluation criteria for clinical TNM staging is shown in Supplementary Material 3. Clinical tumor and metastatic lymph node staging details are summarized in Tables S1 and S2. All patients were followed up every 2–3 months until death or loss to follow-up. The follow-up period was measured from enrollment until May 2021. Information regarding survival status was collected during the last follow-up. Two survival arms, PFS and OS, were recorded and considered as the references for further analysis. PFS was measured from the start of treatment to the date when disease progression was documented. OS was measured from the start of treatment to the date of death. All patients completed all relevant routine tests (Supplementary Material 1). Patients from all three centers underwent CT. The CT protocol and parameters of the three centers are shown in Table S3.

      Image Segmentation

      The open-source software ITK-SNAP (www.itksnap.org) was used for manual segmentation. Two-dimensional regions of interest (ROI) were delineated for the primary tumors and three main metastatic lesions (ROI-T for the primary tumor, ROI-PM for the PM region, ROI-LM for LM, and ROI-LNM for LNM). Details of the segmentation criteria of the four ROIs are presented in Supplementary Material 4. The workflow of this study is shown in Figure 2. Feature extraction was completed using radiomics analysis software on a research platform (Radiomics, Research Frontier, SyngoVia, Version VB20, Siemens Healthineers) in accordance with the image biomarker standardization initiative (
      • Zwanenburg A
      • Vallieres M
      • Abdalah MA
      • et al.
      The image biomarker standardization initiative: standardized quantitative radiomics for high-throughput image-based phenotyping.
      ). Details of the feature extraction process are shown in Supplementary Material 4. In total, 854 radiomic features were extracted from each ROI (Fig. S1). Details of the feature reliability analysis are presented in Supplementary Material 4.
      Figure 2
      Figure 2The workflow of this study. LM, liver metastasis; LNM, lymph node metastasis, NIPS: neoadjuvant intraperitoneal and systemic chemotherapy; PM, peritoneal metastasis; T, primary tumor.

      Generation of Lesion Markers

      For feature screening and marker training, data on all patients in the training cohort were used to construct the primary tumor marker (T_marker), whereas those on 84, 34, and 80 patients with LNM, LM, and PM, respectively, were used to construct the LNM_marker, LM_marker, and PM_marker, respectively. Since not all patients presented all three metastases concurrently, we transformed the three metastasis markers into ordinal categorical variables. In the training cohort, markers in patients with metastasis were stratified into low and high tiers based on the optimal cutoff value determined using the X-tile software tool (
      • Camp RL
      • Dolled-Filhart M
      • Rimm DL.
      X-tile: a new bio-informatics tool for biomarker assessment and outcome-based cut-point optimization.
      ), to which the values one and two were assigned, respectively. If without metastasis, the corresponding lesion marker was recorded as zero. For the feature selection, the Boruta algorithm and Spearman correlation coefficient were calculated to overcome overfitting and to screen out the relevant but nonredundant features of the final training markers. Detailed information on the feature dimensionality reduction is shown in Supplementary Material 5. The random survival forest algorithm with a 10-fold cross-validation parameter adjusted using the grid search approach was implemented for marker generation. Details of the generation of the four lesion markers are shown in Supplementary Material 5.

      Radiomics Model and Radiomic Score

      The radiomics model was constructed using multivariate Cox regression analysis of the T_marker and three metastasis markers to predict PFS. Individual risk and radiomic scores were generated for each patient.
      The prognostic accuracy of the radiomics model for PFS was assessed in the training and two independent validation cohorts using time-dependent receiver operating characteristic curve analysis and Harrell's concordance index (C-index). The goodness of fit of the radiomics model was evaluated using the calibration curves.
      In the training cohort, the optimal cutoff value of the radiomic score for stratifying PFS was determined using X-tile (
      • Camp RL
      • Dolled-Filhart M
      • Rimm DL.
      X-tile: a new bio-informatics tool for biomarker assessment and outcome-based cut-point optimization.
      ). Patients who scored below the cutoff were classified into the low score group, whereas those with higher scores were classified into the high score group. The association of radiomic score with PFS and OS was initially assessed in the training cohort and then validated in the internal and external validation cohorts based on Kaplan–Meier (KM) survival analysis. Furthermore, the stratification ability of the radiomic score of whole cohorts was further verified by subgroup analysis in the three main metastasis subgroups and surgery/nonsurgery subgroups after perioperative chemotherapy.

      Clinical and Comprehensive Models

      Besides the radiomics model, clinical and comprehensive (Rad-clinical model) models were also established to further explore and verify the incremental value of the CT-based radiomic scoring system. The predictive performances and clinical benefits of the three models were then compared in both the internal and external validation cohorts using C-index and decision curve analysis, respectively. Details on the establishment of the clinical and Rad-clinical models are presented in Supplementary Material 6.

      Radiomic Score and Treatment Benefit

      We explored the association between radiomic score and efficacy of NIPS therapy (based on PFS and OS) in the entire cohort, as well as in the PM and non-PM subgroups. Stratified analysis was performed according to the radiomic score level associated with the efficacy of NIPS therapy to identify patient subgroups that could benefit from NIPS therapy. In addition, we also examined the relationship between the radiomic score and surgery during follow-up using the Mann–Whitney U test.

      Statistical Analysis

      All statistical analyses were performed using R (version 3.6.0) and the Python scikit-survival package (version 3.7, scikit-survival package version 0.13.2). Details are described in Supplementary Material 7. Scikit-survival package resources are listed in Table S4. Two-sided p < 0.05 was defined as statistically significant.

      Results

      Patient Characteristics

      The baseline characteristics of the 283 enrolled FGC patients are summarized in Table 1. The median PFS and OS were 9 and 16 months in the training cohort, respectively; 9 and 15 months in the internal validation cohort, respectively; and 17 and 23.5 months in the external validation cohort, respectively. Significant differences in PFS and OS were observed between the training and external validation cohorts (p < 0.05). Sixty-three patients underwent surgery during the follow-up period (training cohort, 24; internal validation cohort, 13; external validation cohort, 26). Patient clinical characteristics are shown in Table S5.
      Table 1Demographic and Clinicopathological Characteristics
      Training cohort (n = 166)Internal validation cohort (n = 83)External validation cohort (n = 34)
      Age (years, median)60 (IQR:50.5–66)60.5 (IQR:49.5–68)63 (IQR:56.5–71)
      Sex (n, %)Male107 (64.5%)49 (59.0%)24 (70.6%)
      Female59 (35.5%)34 (41.0%)10 (29.4%)
      PS score (n, %)06 (3.6%)5 (6.0%)17 (50.0%)
      1157 (94.6%)157 (94.6%)11 (32.4%)
      23 (1.8%)0 (0.0%)6 (17.6%)
      Location (n, %)Proximal42 (25.3%)15 (18.1%)6 (17.6%)
      Middle62 (37.3%)31 (37.3%)12 (35.3%)
      Distal62 (37.3%)37 (44.6%)16 (47.1%)
      Clinical TNM stage (n, %)III31 (18.7%)23 (27.7%)18 (52.9%)
      IV135 (81.3%)60 (72.3%)16 (47.1%)
      Clinical T stage (n, %)T4a123 (74.1%)65 (78.3%)30 (88.2%)
      T4b43 (25.9%)18 (21.7%)4 (11.8%)
      Lymph node status (n, %)Negative82 (49.4%)49 (59.0%)13 (38.2%)
      Positive84 (50.6%)34 (41.0%)21 (61.8%)
      Liver metastasis (n, %)Negative132 (79.5%)57 (68.7%)30 (88.2%)
      Positive34 (20.5%)26 (31.3%)4 (11.8%)
      Peritoneal metastasis (n, %)Negative86 (51.8%)41 (49.4%)28 (82.4%)
      Positive80 (48.2%)42 (50.6%)6 (17.6%)
      Other organ metastasis (n, %)Yes22 (13.3%)15 (18.1%)3 (8.8%)
      No144 (86.7%)68 (81.9%)31 (91.2%)
      CEA (ng/mL, %)≤5107 (64.5%)45 (54.2%)25 (73.5%)
      >559 (35.5%)38 (45.8%)9 (26.5%)
      CA125(U/mL, %≤3596 (57.8%)51 (61.4%)24 (70.6%)
      >3570 (42.2%)32 (38.6%)10 (29.4%)
      NIPS therapy (n, %)Yes28 (16.9%)11 (13.3%)7 (20.6%)
      No138 (83.1%)72 (86.7%)27 (79.4%)
      Surgery (n, %)Yes24 (14.5%)13 (15.7%)26 (76.5%)
      No142 (85.5%)70 (84.3%)8 (23.5%)
      CEA, carcinoembryonic antigen; CA125, carbohydrate antigen 125; IQR, interquartile range; NIPS, neoadjuvant intraperitoneal and systemic; PS score, performance status score; TNM, tumor node-metastasis.

      Lesion Markers

      In all, 692, 623, 656, and 674 radiomic features showed robustness with an intraclass coefficient (ICC) higher than 0.8 for the primary tumor, PM, LM, and LNM images, respectively. Details of the intraobserver ICC results are shown in Supplementary Material 8. Only 13, 9, 9, and 8 important features were identified through lesion marker construction of the primary tumor, PM, LM, and LNM images, respectively (Fig. S2). Detailed information on the results of the three metastasis marker constructions is shown in Supplementary Material 8. The validation of the markers is demonstrated in Figure S3.

      Radiomics Model and Radiomic Score

      The radiomics model was composed of the T_markers and three metastasis markers. For PFS, the T_marker, PM_marker, and LM_marker were identified as risk factors, with hazard ratios (HR) of 1.14 (95% confidence interval [CI]: 1.10–1.20), 4.17 (95% CI: 2.00–8.70), and 4.28 (95% CI: 2.31–7.90), respectively (all p < 0.001). The formula of radiomic score is shown in Supplementary Material 9. Detailed information is shown in Figure S4.
      The radiomics model had a C-index of 0.83 in the training cohort, 0.75 in the internal validation cohort, and 0.76 in the external validation cohort (Table 2). The time-dependent cumulative dynamic area under the curve (AUC) in both the training and validation cohorts showed high prognostic accuracy (average AUC 0.88 and 0.83, respectively; Fig. S5). The calibration curves for the 1-year and 2-year PFS indicated a high goodness of fit of the radiomics model. However, the 2-year PFS calibration curve showed a relatively poor consistency in the external validation cohort (Fig. S6).
      Table 2The Performances of the Different Models in the Training and Validation Cohorts
      CohortC-index (95% CI)
      Radiomics modelClinical modelRad-clinical model
      Training cohort0.83
      p <0.05, compared with clinical model.
      (0.788–0.872)
      0.64 (0.580–0.700)0.83
      p <0.05, compared with clinical model.
      (0.789–0.871)
      Internal validation cohort0.75
      p <0.05, compared with clinical model.
      (0.682–0.818)
      0.54 (0.459–0.621)0.75
      p <0.05, compared with clinical model.
      (0.693–0.827)
      External validation cohort0.76 (0.669–0.851)0.60 (0.452–0.748)0.78
      p <0.05, compared with clinical model.
      (0.688–0.872)
      CI, confidence interval
      low asterisk p <0.05, compared with clinical model.
      The cutoff value of the radiomic score in the training cohort was 1.06. The risk stratification ability of the radiomic score for the PFS and OS was well verified in both the training and internal validation cohorts (all p < 0.05; Fig. 3). In the external validation cohort, the risk stratification ability of the radiomic score was well verified only for PFS (p < 0.05; Fig. 3). These results suggest that patients with FGC with a high radiomic score have worse outcomes (median PFS: 5 months vs. 16 months; median OS: 18 months vs. >40 months; all p < 0.001) (Table 3).
      Figure 3
      Figure 3K-M curves of progression-free survival and overall survival according to radiomic score among overall far-advanced gastric cancer patients. (a). Training cohort. (b). Internal validation cohort. (c). External validation cohort. p values were calculated using a two-sided log-rank test. KM curve, Kaplan–Meier curve.
      Table 3NIPS Therapy Interaction with Radiomic Score for PFS and OS in Patients with far-Advanced Gastric Cancer
      Number of patientsMedian PFS months (95%CI)Median OS months (95%CI)Number of patientsMedian PFS months (95%CI)Median OS months (95%CI)p value (PFS)p value (OS)
      High score groupLow score group
      Overall cohorts635 (5–6)18 (17–NA)22016 (14–19)>40 (40–NA)<0.001
      p < 0.05.
      <0.001
      p < 0.05.
      Peritoneal metastasis315 (5–7)17 (11–NA)9714 (11–18)33 (26–NA)<0.001
      p < 0.05.
      <0.001
      p < 0.05.
      NIPSNon-NIPS
      Peritoneal metastasis2614 (8–NA)33 (25–NA)10210 (8–13)26 (20–NA)0.260<0.001
      p < 0.05.
      High score group1110.7 (7–NA)>40(–)627 (6–9)20 (15–NA)0.1600.043
      p < 0.05.
      Low score group1517 (8–NA)33 (25–NA)4021 (16–NA)>40(26–NA)0.5300.580
      CI, confidence interval; NIPS, neoadjuvant intraperitoneal and systemic; OS, overall survival; PFS, progression-free survival.
      low asterisk p < 0.05.
      For risk stratification in the various cohorts, Kaplan–Meier survival curves were plotted for eight subgroups based on the previous cutoff value of 1.06. In the eight subgroups, patients could be classified into low or high PFS or OS (all p < 0.05) using the radiomic score (Fig. 4). Specifically, in the PM subgroup, patients in the low score group had significantly longer PFS and OS than those in the high score group (median PFS: 14 months vs. 5 months; median OS: 33 months vs. 17 months; all p < 0.001) (Table 3).
      Figure 4
      Figure 4K-M curves of progression-free survival and overall survival according to radiomic score in the main metastasis and surgery subgroups. (a). The liver metastasis (LM) subgroup. (b). Non-LM subgroup. (c). LN metastasis (LNM) subgroup. (d). Non-LNM subgroup. (e). The peritoneal metastasis (PM) subgroup. (f). Non-PM subgroup. (g). Surgery subgroup. (h). Non-surgery subgroup. p values were calculated using a two-sided log-rank test. KM curve, Kaplan–Meier curve.

      Clinical and Rad-clinical Models

      Univariate and multivariate Cox regression analyses identified PS score, cT, clinical distant metastasis (cM) stage, and CEA as significant prognostic factors for constructing a clinical model (Fig. S7), indicating a poor ability to predict prognosis (training cohort: C-index 0.64, 95% CI: 0.580–0.700; internal validation cohort: C-index 0.54, 95% CI: 0.459–0.621; external validation cohort: C-index 0.60, 95% CI: 0.452–0.748) (Table 2). After stepwise model selection, the radiomic score (HR: 2.80, 95% CI: 2.30–3.30; p < 0.05) and CA125 (HR: 1.60, 95% CI: 1.10–2.40; p < 0.05) were identified as independent risk factors for PFS (Fig. S8).
      Compared with the clinical model, the radiomics model significantly improved the prognostic performance (training cohort: C-index 0.83, 95% CI: 0.788–0.872; internal validation cohort: C-index 0.75, 95% CI: 0.682–0.818; external validation cohort: C-index 0.76, 95% CI: 0.669–0.851; all p < 0.05 compared with the clinical model) (Table 2). However, no significant differences were found between the radiomics model and the Rad-clinical model across all three cohorts (all p > 0.05; Table 2). The decision curves (DCA) indicated that the radiomics and Rad-clinical models were both more beneficial in shaping treatment strategies than the clinical model across the whole threshold range (Fig. 5), which further proved the clinical benefit of the radiomics model.
      Figure 5
      Figure 5DCA curves for the radiomics model, clinical model and Rad-clinical model. DCA, decision curve analysis.

      Radiomic Score and Treatment Benefit

      The interaction test for radiomic score and efficacy of NIPS therapy revealed a higher benefit from NIPS among the high radiomic score group with PM than among the low-radiomic score group. In the high score group with PM, the use of NIPS treatment prolonged the OS, and a significant difference in the prognosis was observed between the NIPS and non-NIPS groups (median OS: >40 months vs. 20 months; p = 0.043). In the low score group with PM, no obvious superiority in PFS and OS were observed between patients with or without NIPS therapy (median PFS: 17 months vs. 21 months; median OS: 33 months vs. >40 months; all p > 0.05) (Fig. 6 and Table 3).
      Figure 6
      Figure 6NIPS benefits based on K-M curves of progression-free survival and overall survival according to radiomic score among far-advanced gastric cancer patients. (a). High score peritoneal metastasis (PM) subgroup. (b). Low score PM subgroup. (c). High score non-PM subgroup. (d). Low score non-PM subgroup. p values were calculated using a two-sided log-rank test. KM curve, Kaplan–Meier curve; NIPS, neoadjuvant intraperitoneal and systemic chemotherapy.
      A significant difference was found in both PFS and OS status between the surgery and non-surgery subgroups (p < 0.05; Fig. 4G–4H). In terms of the relationship between radiomic score and surgery, patients with low radiomic scores were more likely to be indicated for surgery after perioperative chemotherapy (p < 0.05). Of the entire study population, 95.2% (60/63) of patients who were stratified into the low score group underwent surgery and obtained an OS > 40 months (Table S6).

      DISCUSSION

      In the present study, radiomic scores were determined based on the CT imaging features of both the primary tumor and main metastatic lesions and subsequently validated not only to predict the prognosis in FGC, but also to stratify survival in FGC in all three cohorts, metastasis subgroups, and in the surgery and nonsurgery subgroups. We found that the high score group had poor OS, whereas the low score group had better OS. Patients with a low radiomic score were more likely to undergo surgery after perioperative chemotherapy. More importantly, we further identified that patients with PM and a high radiomic score would benefit from NIPS therapy.
      A significant difference in the prognosis was noted between high risk and low risk patients grouped by radiomic analysis in tumors other than FGC (
      • Liu Z
      • Meng X
      • Zhang H
      • et al.
      Predicting distant metastasis and chemotherapy benefit in locally advanced rectal cancer.
      ,
      • Meng Y
      • Zhang Y
      • Dong D
      • et al.
      Novel radiomic signature as a prognostic biomarker for locally advanced rectal cancer.
      ). In our study, we demonstrated the good ability of the radiomic score to predict PFS and OS, as well as to stratify PFS and OS in FGC (median PFS: 5 months vs. 16 months; median OS: 18 months vs. >40 months; all p < 0.001). Furthermore, the radiomic score can also be used to identify patients with different PFS and OS in the same metastasis subgroups in both the training and validation cohorts, which has not been reported previously. Patients with FGC who successfully underwent surgery after downstaging by chemotherapy have more favorable long-term survival rates than those undergoing chemotherapy alone. Consistent with previous reports (
      • Gong Y
      • Wang P
      • Zhu Z
      • et al.
      Benefits of Surgery After Neoadjuvant Intraperitoneal and Systemic Chemotherapy for Gastric Cancer Patients With Peritoneal Metastasis: A Meta-Analysis.
      ), of the 63 patients who underwent surgery in the present study, 60 (95.2%) were in the low score group and had a median survival of more than 40 months. The radiomic score also guides the personalized treatment of patients, even those in the same metastasis subgroups. This might be because our radiomics model contains primary tumor and main metastasis features to which prognosis is closely related, including LM, PM, and LNM. Considering the comprehensive analysis of characteristics of patients with FGC patients, the stratification rate appears to be better, as confirmed in the external validation cohort.
      Another important clinical significance of the radiomic score is in its value in stratifying the survival risk in FGC and identifying patients indicated for NIPS therapy, which can improve the successful radical resection and OS rates. Although the current guidelines recommend NIPS therapy for patients with FGC with PM, some studies have found that not all patients could benefit from the therapy, which may be due to the scope and heterogeneity of PM (
      • Yamaguchi H
      • Kitayama J
      • Ishigami H
      • et al.
      A phase 2 trial of intravenous and intraperitoneal paclitaxel combined with S-1 for treatment of gastric cancer with macroscopic peritoneal metastasis.
      ). In our study, patients with FGC with high radiomic scores in the PM subgroup had significantly higher OS after NIPS therapy (median OS: >40 months vs. 20 months; p = 0.043). In addition, patients with FGC who underwent NIPS therapy may not show a better prognosis than those who did not and had a low radiomic score (median OS: 33 months vs. >40 months; p > 0.05). Thus, the radiomic score may be useful for identifying patients with FGC with PM who can benefit from NIPS therapy.
      We also established a comprehensive Rad-clinical model that was equally effective in predicting survival, indicating that clinical features achieve no improvement to the existing result. The tumor markers CEA, CA199, and CA724 have been highly useful for staging before surgery or after chemotherapy and especially for monitoring recurrence or evaluating treatment response in advanced gastric cancer (
      • Emoto S
      • Ishigami H
      • Yamashita H
      • et al.
      Clinical significance of CA125 and CA72-4 in gastric cancer with peritoneal dissemination.
      ). In our study, we focused on prognostic stratification before treatment using only baseline imaging and tumor indicators. Furthermore, the radiomic score and CA125 level were identified as independent prognostic clinical factors for FGC. CA125 level has been associated with peritoneal dissemination in gastric cancer (
      • Yang C
      • Yang Y
      • Huang X
      • et al.
      A nomogram based on clinicopathologic features and preoperative hematology parameters to predict occult peritoneal metastasis of gastric cancer: a single-center retrospective study.
      ,
      • Shimada H
      • Noie T
      • Ohashi M
      • et al.
      Clinical significance of serum tumor markers for gastric cancer: a systematic review of literature by the task force of the Japanese Gastric Cancer Association.
      ). In our study, the positivity for CA125 was up to 42.3%, which was likely reflected in the comprehensive model. The CT-based radiomic score established in this study thus addresses the lack of a straightforward and noninvasive method for predicting prognosis and stratifying survival in FGC before treatment and provides additional value as an imaging prognostic marker for clinicians.
      Although this was a multicenter study, it has some limitations. First, the sample size is limited, and the data were collected retrospectively, which requires further larger, well-designed prospective studies before clinical application. The accumulation of additional patients, especially in the external validation cohort, will diversify the patient- and tumor-specific information collected and help construct a more stable and accurate model. Second, while imaging features focus on the primary and main metastatic information from lesions, the inclusion of data on rare metastatic sites, such as lung or bone, and genomic sequencing may generate more data needed to increase the performance of models. Third, we only examined the performance of the radiomic score in a Chinese population, and its prognostic ability in different ethnic populations is unclear, thereby further necessitating future studies.
      In conclusion, we used a CT-based radiomic score to effectively predict PFS and OS and stratify the risk of survival in FGC, which helps create opportunities for early intervention, such as surgery and improve outcomes. In addition, the radiomic score may be a useful imaging tool for selecting patients who can benefit from NIPS therapy.

      Funding

      This work was supported by the National Science Foundation of China (81972707; 82271934; 82273126; 81771789) and Collaborative innovation cluster project of Shanghai Municipal Health Commission (2020CXJQ03).

      Appendix. Supplementary materials

      References

        • Siegel RL
        • Miller KD
        • Fuchs HE
        • et al.
        Cancer statistics, 2021.
        CA Cancer J Clin. 2021; 71: 7-33
        • Cheng XJ
        • Lin JC
        • Tu SP.
        Etiology and prevention of gastric cancer.
        Gastrointest Tumors. 2016; 3: 25-36
        • Saito M
        • Kiyozaki H
        • Takata O
        • et al.
        Treatment of stage IV gastric cancer with induction chemotherapy using S-1 and cisplatin followed by curative resection in selected patients.
        World J Surg Oncol. 2014; 12: 406
        • Uemura N
        • Kikuchi S
        • Sato Y
        • et al.
        A phase II study of modified docetaxel, cisplatin, and S-1 (mDCS) chemotherapy for unresectable advanced gastric cancer.
        Cancer Chemother Pharmacol. 2017; 80: 707-713
        • Bilici A
        • Selcukbiricik F.
        Prognostic significance of the recurrence pattern and risk factors for recurrence in patients with proximal gastric cancer who underwent curative gastrectomy.
        Tumour Biol. 2015; 36: 6191-6199
        • Nakagawa M
        • Kojima K
        • Inokuchi M
        • et al.
        Patterns, timing and risk factors of recurrence of gastric cancer after laparoscopic gastrectomy: reliable results following long-term follow-up.
        Eur J Surg Oncol. 2014; 40: 1376-1382
        • Macri A
        • Morabito F.
        The use of intraperitoneal chemotherapy for gastric malignancies.
        Expert Rev Anticancer Ther. 2019; 19: 879-888
        • Fujitani K
        • Yang HK
        • Kurokawa Y
        • et al.
        Randomized controlled trial comparing gastrectomy plus chemotherapy with chemotherapy alone in advanced gastric cancer with a single non-curable factor: Japan Clinical Oncology Group Study JCOG 0705 and Korea Gastric Cancer Association Study KGCA01.
        Jpn J Clin Oncol. 2008; 38: 504-506
        • Yamaguchi H
        • Kitayama J
        • Ishigami H
        • et al.
        A phase 2 trial of intravenous and intraperitoneal paclitaxel combined with S-1 for treatment of gastric cancer with macroscopic peritoneal metastasis.
        Cancer. 2013; 119: 3354-3358
        • Ishigami H
        • Kitayama J
        • Kaisaki S
        • et al.
        Phase II study of weekly intravenous and intraperitoneal paclitaxel combined with S-1 for advanced gastric cancer with peritoneal metastasis.
        Ann Oncol. 2010; 21: 67-70
        • Gong Y
        • Wang P
        • Zhu Z
        • et al.
        Benefits of Surgery After Neoadjuvant Intraperitoneal and Systemic Chemotherapy for Gastric Cancer Patients With Peritoneal Metastasis: A Meta-Analysis.
        J Surg Res. 2020; 245: 234-243
        • Ishigami H
        • Fujiwara Y
        • Fukushima R
        • et al.
        Phase III Trial Comparing Intraperitoneal and Intravenous Paclitaxel Plus S-1 Versus Cisplatin Plus S-1 in Patients With Gastric Cancer With Peritoneal Metastasis: PHOENIX-GC Trial.
        J Clin Oncol. 2018; 36: 1922-1929
        • Wang S
        • Feng C
        • Dong D
        • et al.
        Preoperative computed tomography-guided disease-free survival prediction in gastric cancer: a multicenter radiomics study.
        Med Phys. 2020; 47: 4862-4871
        • Nakamoto T
        • Haga A
        • Takahashi W.
        [An Introduction to Radiomics: Toward a New Era of Precision Medicine].
        Igaku Butsuri. 2018; 38: 129-134
        • Zhang L
        • Dong D
        • Zhang W
        • et al.
        A deep learning risk prediction model for overall survival in patients with gastric cancer: A multicenter study.
        Radiother Oncol. 2020; 150: 73-80
        • Sun KY
        • Hu HT
        • Chen SL
        • et al.
        CT-based radiomics scores predict response to neoadjuvant chemotherapy and survival in patients with gastric cancer.
        BMC Cancer. 2020; 20: 468
        • Xie W
        • Jiang Z
        • Zhou X
        • et al.
        Quantitative Radiological Features and Deep Learning for the Non-Invasive Evaluation of Programmed Death Ligand 1 Expression Levels in Gastric Cancer Patients: A Digital Bopsy Study.
        Acad Radiol. 2022;
        • Liu S
        • Qiao X
        • Xu M
        • et al.
        Development and validation of multivariate models integrating preoperative clinicopathological parameters and radiographic findings based on late arterial phase CT images for predicting lymph node metastasis in gastric cancer.
        Acad Radiol. 2021; 28 (Suppl): S167-SS78
        • Liao F
        • Guo X
        • Lu X
        • Dong W.
        A validated survival nomogram for early-onset diffuse gastric cancer.
        Aging (Albany NY). 2020; 12: 13160-13171
        • Shin J
        • Lim JS
        • Huh YM
        • et al.
        A radiomics-based model for predicting prognosis of locally advanced gastric cancer in the preoperative setting.
        Sci Rep. 2021; 11: 1879
        • Li W
        • Zhang L
        • Tian C
        • et al.
        Prognostic value of computed tomography radiomics features in patients with gastric cancer following curative resection.
        Eur Radiol. 2019; 29: 3079-3089
        • Wei XL
        • Xu JY
        • Wang DS
        • et al.
        Baseline lesion number as an efficacy predictive and independent prognostic factor and its joint utility with TMB for PD-1 inhibitor treatment in advanced gastric cancer.
        Ther Adv Med Oncol. 2021; 131758835921988996
        • Granieri S
        • Altomare M
        • Bruno F
        • et al.
        Surgical treatment of gastric cancer liver metastases: Systematic review and meta-analysis of long-term outcomes and prognostic factors.
        Crit Rev Oncol Hematol. 2021; 163103313
        • Ter Veer E
        • van Kleef JJ
        • Schokker S
        • et al.
        Prognostic and predictive factors for overall survival in metastatic oesophagogastric cancer: A systematic review and meta-analysis.
        Eur J Cancer. 2018; 103: 214-226
        • Wang FH
        • Zhang XT
        • Li YF
        • et al.
        The Chinese Society of Clinical Oncology (CSCO): clinical guidelines for the diagnosis and treatment of gastric cancer, 2021.
        Cancer Commun (Lond). 2021; 41: 747-795
        • Sano T
        • Coit DG
        • Kim HH
        • et al.
        Proposal of a new stage grouping of gastric cancer for TNM classification: International Gastric Cancer Association staging project.
        Gastric Cancer. 2017; 20: 217-225
        • Zwanenburg A
        • Vallieres M
        • Abdalah MA
        • et al.
        The image biomarker standardization initiative: standardized quantitative radiomics for high-throughput image-based phenotyping.
        Radiology. 2020; 295: 328-338
        • 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: 7252-7259
        • Liu Z
        • Meng X
        • Zhang H
        • et al.
        Predicting distant metastasis and chemotherapy benefit in locally advanced rectal cancer.
        Nat Commun. 2020; 11: 4308
        • Meng Y
        • Zhang Y
        • Dong D
        • et al.
        Novel radiomic signature as a prognostic biomarker for locally advanced rectal cancer.
        J Magn Reson Imaging. 2018;
        • Emoto S
        • Ishigami H
        • Yamashita H
        • et al.
        Clinical significance of CA125 and CA72-4 in gastric cancer with peritoneal dissemination.
        Gastric Cancer. 2012; 15: 154-161
        • Yang C
        • Yang Y
        • Huang X
        • et al.
        A nomogram based on clinicopathologic features and preoperative hematology parameters to predict occult peritoneal metastasis of gastric cancer: a single-center retrospective study.
        Dis Markers. 2020; 20201418978
        • Shimada H
        • Noie T
        • Ohashi M
        • et al.
        Clinical significance of serum tumor markers for gastric cancer: a systematic review of literature by the task force of the Japanese Gastric Cancer Association.
        Gastric Cancer. 2014; 17: 26-33