Deep Learning Model Based on Tumor and Visceral Adipose Tissue CT Features for Predicting Peritoneal Metastasis Risk after Radical Gastrectomy in Serosa-Invasive Gastric Cancer.

Deep Learning Model Based on Tumor and Visceral Adipose Tissue CT Features for Predicting Peritoneal Metastasis Risk after Radical Gastrectomy in Serosa-Invasive Gastric Cancer - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Radiol Imaging Cancer . 2026 Apr 24;8(3):e250353. doi: 10.1148/rycan.250353 Deep Learning Model Based on Tumor and Visceral Adipose Tissue CT Features for Predicting Peritoneal Metastasis Risk after Radical Gastrectomy in Serosa-Invasive Gastric Cancer Yueyue Li 1 , Ximiao Wang 2 , Qiuying Chen 1 , Hua Xiao 3 , Yilong Huang 4, 5, 6 , Liebin Huang 7 , Lian Jian 8 , Wansheng Long 2, 7 , Bao Feng 2, 9 , Shuixing Zhang 1, ✉ Author information Article notes Copyright and License information 1 Department of Radiology, The First Affiliated Hospital of Jinan University, No. 613 Huangpu Avenue West, Tianhe District, Guangzhou 510632, People’s Republic of China 2 Jiangmen Key Laboratory of Artificial Intelligence in Medical Image Computation and Application, Jiangmen Central Hospital, Jiangmen, People’s Republic of China 3 Department of Hepatobiliary and Intestinal Surgery, Hunan Cancer Hospital, The Affiliated Cancer Hospital of Xiangya School of Medicine, Central South University, Changsha, Hunan, People’s Republic of China 4 Department of Medical Imaging, The First Affiliated Hospital of Kunming Medical University, Kunming, People’s Republic of China 5 Department of Radiology, Guangdong Provincial People’s Hospital (Guangdong Academy of Medical Sciences), Southern Medical University, Guangzhou, People’s Republic of China 6 Guangdong Provincial Key Laboratory of Artificial Intelligence in Medical Image Analysis and Application, Guangzhou, People’s Republic of China 7 Department of Radiology, Jiangmen Central Hospital, Jiangmen, People’s Republic of China 8 Department of Radiology, Hunan Cancer Hospital, The Affiliated Cancer Hospital of Xiangya School of Medicine, Central South University, Changsha, People’s Republic of China 9 Laboratory of Intelligent Detection and Information Processing, Guilin University of Aerospace Technology, Guilin, People’s Republic of China ✉ Address correspondence to: S.Z. (email: shui7515@126.com ). ✉ Corresponding author. Roles Yueyue Li : MD , Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Validation, Visualization, Writing - original draft, Writing - review & editing Ximiao Wang : PhD , Conceptualization, Formal analysis, Investigation, Methodology, Software, Validation, Visualization, Writing - original draft, Writing - review & editing Qiuying Chen : MD , Data curation, Formal analysis, Methodology, Validation, Visualization, Writing - original draft, Writing - review & editing Hua Xiao : MD , Data curation, Formal analysis, Validation, Visualization Yilong Huang : PhD , Formal analysis, Investigation, Methodology, Software, Validation, Visualization, Writing - original draft, Writing - review & editing Liebin Huang : MD , Data curation, Methodology, Validation Lian Jian : MD , Data curation, Methodology, Validation Wansheng Long : MD , Project administration, Supervision, Writing - review & editing Bao Feng : PhD , Software, Supervision, Validation Shuixing Zhang : MD, PhD , Funding acquisition, Project administration, Supervision, Writing - review & editing Received 2025 Jul 5; Revision requested 2025 Aug 1; Revised 2026 Feb 6; Accepted 2026 Mar 6; Collection date 2026 May. © 2026 by the Radiological Society of North America, Inc. PMC Copyright notice PMCID: PMC13231214  PMID: 42029163 Abstract Purpose To develop and validate a deep learning model integrating tumor and visceral adipose tissue (VAT) CT scan features with clinical indicators to predict postoperative peritoneal metastasis in serosa-invasive gastric cancer. Materials and Methods This multicenter, retrospective study between April 2008 and January 2018 included patients with pathologically confirmed serosa-invasive gastric cancer. Patients were divided into training, internal test, and independent external test sets. Tumor and VAT regions were segmented at preoperative CT. Deep features were extracted using a ResNet18 network. A fused tumor-VAT deep learning signature (F-DLS) was generated, incorporating clinical variables into a multimodal deep learning radiomics model (MDLR) using a sparse Bayesian extreme learning machine. Model performance was assessed using receiver operating characteristic curve, integrated discrimination improvement, calibration, decision curve analysis, and recurrence-free survival. Results Among 416 patients (mean age, 56.6 years ± 11.6; 66.1% male patients), the F-DLS achieved area under the receiver operating characteristic curve (AUC) values of 0.81 (95% CI: 0.73, 0.88) in the internal test set and 0.79 (95% CI: 0.71, 0.86) in the external test set. Compared with the tumor tissue DLS and VAT-DLS, the F-DLS showed numerically higher AUCs without statistical significance. The MDLR achieved the strongest predictive performance, with AUCs of 0.86 (95% CI: 0.79, 0.92) in the internal test set and 0.86 (95% CI: 0.78, 0.92) in the external test set. The MDLR statistically significantly outperformed clinical and deep learning–only models (integrated discrimination improvement, P < .001), showed good calibration, and provided favorable net benefit on decision curve analysis. High-risk patients identified by the MDLR had significantly shorter recurrence-free survival (log-rank P < .001). Conclusion The MDLR integrating CT scan features and clinical indicators enabled noninvasive prediction of peritoneal metastasis risk in serosa-invasive gastric cancer and may facilitate postoperative risk stratification. Keywords: Gastric Cancer, Peritoneal Metastasis, CT, Visceral Adipose Tissue, Deep Learning Supplemental material is available for this article. © RSNA, 2026 Keywords: Gastric Cancer, Peritoneal Metastasis, CT, Visceral Adipose Tissue, Deep Learning Open in a new tab Summary A deep learning model integrating tumor and visceral adipose tissue features on CT images with clinical indicators showed strong performance in predicting postoperative peritoneal metastasis in patients with serosa-invasive gastric cancer. Key Points ■ In this retrospective study, a multimodal deep learning radiomics model (MDLR) to predict peritoneal metastasis after radical gastrectomy was developed by combining tumor and visceral adipose tissue CT scan features with clinical data from 416 patients with serosa-invasive gastric cancer. ■ The model showed strong performance, with area under the receiver operating characteristic curve values of 0.86 (95% CI: 0.79, 0.92) in the internal test set and 0.86 (95% CI: 0.78, 0.92) in the external test set. ■ Compared with low-risk patients, high-risk patients identified by the MDLR had shorter recurrence-free survival (mean, 30.0 months ± 27.3 vs 46.1 months ± 32.0; log-rank P < .001). Introduction Gastric cancer (GC) is the fifth most common cancer and the third leading cause of cancer-related deaths worldwide ( 1 ). Peritoneal metastasis (PM) is the most common metastatic pattern in GC, with up to 50% of patients with stage T3 or T4 tumors developing PM after radical surgery ( 2 – 4 ). The outcomes for patients with PM are extremely poor: the median survival is only 3–6 months, and the 5-year overall survival rate is less than 2% ( 5 ). GC serosal invasion is a known high-risk factor for PM ( 6 ). Even after curative resection, up to 50% of patients with serosal invasion develop peritoneal recurrence ( 5 ). Therefore, it is important to identify high-risk patients early and accurately. To reduce the risk of PM, intraperitoneal chemotherapy (IPC) has been used either during or shortly after surgery ( 7 ). In IPC administration, chemotherapeutic agents are directly delivered into the peritoneal cavity, which increases the local concentration of these agents to kill residual tumor cells. Some study findings suggest that IPC may improve outcomes in select patients with GC ( 8 , 9 ). However, IPC is costly, increases the risk of postoperative complications, and is not needed for all patients ( 10 ). As such, a reliable method to predict PM risk preoperatively could help identify those patients who could truly benefit from IPC. Currently, PM risk prediction is difficult. Clinical staging systems such as TNM staging provide limited ability to identify patients at high risk, as patients with the same stage of cancer can have markedly different risks of PM development. This limitation is particularly evident in serosa-invasive GC, where all tumors are classified as locally advanced, and thus affected patients cannot be further stratified according to existing staging criteria. Currently used clinical diagnostic approaches, such as CT and peritoneal lavage cytologic analysis, also have some limitations. For example, CT, the most widely used method for detecting PM, can identify only advanced PM with typical features such as omental cake or peritoneal thickening. Cytological analysis of samples obtained from peritoneal lavage may help, but the sensitivity is low, and many patients with negative results still develop PM ( 11 ). These limitations highlight the need for new tools that can assess PM risk in patients with GC with serosal invasion. PM is a complex process involving tumor cell detachment, immune evasion, and implantation at distant peritoneal sites ( 12 , 13 ). In recent years, increasing attention has been given to the tumor microenvironment and host factors involved in this process. Among these, visceral adipose tissue (VAT), which lies close to the stomach and has immune-modulatory properties, may influence peritoneal dissemination of tumor cells through mechanisms such as local inflammation, secretion of adipokines, and support of a protumoral microenvironment ( 14 ). Findings from several studies have linked VAT with chemotherapy response, complications, and survival in GC ( 15 – 17 ); and a higher VAT area has also been reported as a predictor of PM ( 15 , 18 ). These findings suggest that VAT is not simply an energy store but also could actively participate in peritoneal tumor cell dissemination through metabolic or immune-related interactions. CT-based changes in VAT may capture some of this information noninvasively. Deep learning has emerged as a promising tool for extracting image features that are difficult for the human eye to detect. Prior work has shown that deep learning models can predict chemotherapy response, survival outcomes, and even occult PM in GC ( 19 , 20 ). A model combining tumor and VAT features from CT images, enhanced by deep learning, could offer a more complete view of metastatic risk, thereby supporting clinical decision-making, such as whether to administer IPC, and helping avoid unnecessary treatments in patients at low risk. In this study, we aimed to develop and validate a deep learning–based model that integrates CT scan features from the tumor and VAT with clinical indicators to predict PM in patients with serosal invasion who underwent radical gastrectomy. Materials and Methods Study Patients and Treatment This retrospective study was approved by the respective ethical committees of the First Affiliated Hospital of Jinan University and Jiangmen Central Hospital. The requirement to obtain written informed consent was waived owing to the study’s retrospective design. All patient data were anonymized before analysis to ensure confidentiality and compliance with the Declaration of Helsinki. The study adhered to the Standards for Reporting of Diagnostic Accuracy Studies guidelines, where applicable. No industry support was received for this study. Authors had full control of the data and the information submitted for publication. Some patients in this study may have been included in prior publications from the authors’ institutions; however, the study cohort, research question, and analytic strategy of the present study are distinct, and the current analysis has not been previously reported. A total of 416 patients who underwent curative radical gastrectomy were retrospectively selected from two centers. Of these, 302 consecutive patients from center 1 between November 2010 and January 2015 were included and randomly assigned to the training set ( n = 181) or internal test set ( n = 121) in a 6:4 ratio. Patients were included if they (a) underwent radical gastrectomy; (b) had pathologically confirmed GC with serosal invasion; (c) underwent preoperative contrast-enhanced abdominal CT within 4 weeks before surgery, with images available in the picture archiving and communication system; and (d) had available clinicopathology data and postoperative follow-up. Patients were excluded if they had preoperative PM or had received neoadjuvant chemoradiotherapy; the number of excluded patients for each criterion is detailed in the study flowchart ( Fig 1 ). The external test set consisted of 114 consecutive patients from center 2 between April 2008 and January 2018, who met the same inclusion and exclusion criteria as those applied to patients from center 1. The overall test set was defined as the combination of the internal test and external test sets. Figure 1: Open in a new tab Patient inclusion and exclusion flowchart for the study population. GC = gastric cancer, PACS = picture archiving and communication system, PM = peritoneal metastasis. The type of gastrectomy (proximal, distal, or total) and D2 lymphadenectomy performed was based on the Chinese Society of Clinical Oncology Gastric Cancer Guidelines ( 4 ). Either laparoscopic surgery or open surgery was performed depending on tumor location and surgical judgment. Postoperative adjuvant chemotherapy was administered at the discretion of the treating physician unless contraindicated. Chemotherapy regimens included fluoropyrimidine-based monotherapy or combination therapy with fluoropyrimidine and platinum agents. None of the patients in this study received intraoperative or postoperative IPC, ensuring that the occurrence of PM during follow-up was not confounded by prophylactic IPC. Information on baseline clinicopathology variables, including sex, age, tumor location, carcinoembryonic antigen level, differentiation, N stage, TNM stage, positive lymph node ratio, and record of postoperative chemotherapy, were obtained from the electronic medical records. Only variables available before surgery were included to ensure the preoperative applicability of the proposed models. Follow-up and Diagnosis of PM All patients underwent routine follow-up after radical gastrectomy. Follow-up was conducted every 3–6 months during the first 1–2 years, every 6–12 months during years 3–5, and annually thereafter. Routine assessments included laboratory tests (eg, complete blood count, serum tumor marker tests), contrast-enhanced abdominal CT examinations, and endoscopic examinations. During the first 2 years, contrast-enhanced abdominal CT was performed every 6–12 months. When PM was suspected (eg, ascites, abdominal mass), additional abdominal US, contrast-enhanced CT, or PET were performed. PM diagnosis was based on findings from US, CT, or PET imaging during follow-up. Recurrence-free survival (RFS) was defined as the time from radical gastrectomy to first recurrence. Image Acquisition and Processing All patients underwent contrast-enhanced abdominal CT within 4 weeks before surgery. Preoperative contrast-enhanced CT examinations were performed with multidetector CT scanners (GE HealthCare, Siemens Healthineers, and Canon Medical Systems) at the two participating centers. Arterial and portal venous phase images were acquired after intravenous injection of iodinated contrast material. Detailed CT acquisition parameters are provided in Appendix S1 . Venous phase CT images were retrieved from the picture archiving and communication system. Primary tumors were manually segmented by two radiologists (Q.C. and L.J., each with 5 years of experience in abdominal CT) working in consensus and using ITK-SNAP software (version 3.8.0; https://itksnap.org ). Discrepancies were resolved by a senior radiologist (L.H., with 10 years of experience in abdominal CT). All readers were blinded to clinical outcomes and follow-up information at the time of image analysis. VAT was automatically segmented with AutoPanoM (version 1.0.0.0; MediaLab), an in-house software body composition analysis (see Appendix S2 for details). The source code of the network has been publicly released at https://github.com/pupuchuan/BC_Radiomics (commit: aceeba1). To ensure anatomic correspondence between the tumor and VAT data, the VAT segmentation was restricted to the same CT sections where the tumor was present. To ensure segmentation accuracy, all automatically segmented VAT regions were reviewed and manually corrected as needed by a radiologist (Y.H., with 10 years of experience in abdominal CT) using ITK-SNAP. This consensus-based segmentation workflow was designed to minimize interrater variability and ensure consistent region-of-interest delineation. All images of regions of interest were resized to 512 × 512 pixels to ensure consistency. The study workflow is illustrated in Figure 2A . Figure 2: Open in a new tab Workflow of data preprocessing, deep feature extraction, and model construction. (A) Data processing pipeline. (B) Construction of deep learning signatures (DLS) based on tumor and visceral adipose tissue features with ResNet18. (C) Development of a clinical model with sparse Bayesian extreme learning machine (SBELM). (D) Integration of clinical indicators and fused deep learning features into a multimodal deep learning radiomics model (MDLR). CM = clinical model, Conv = convolutional, Fc = full connection, HU = Hounsfield units, ISBELM = integrated SBELM, mRMR = maximum relevance and minimum redundancy, PM = peritoneal metastasis, ROI = region of interest. Deep Learning Signature Construction We fine-tuned two parallel ResNet-18 models, both pretrained on ImageNet, using our dataset to extract deep features from CT images—one for the tumor and one for VAT. We conducted model training with fixed hyperparameters (learning rate of 1e-4, momentum of 0.9, weight decay of 1e-6, batch size of 32, with 150 epochs) and supervised by cross-entropy loss. Parameter training details are provided in Appendix S3 . After training, we used the convolutional filters of the ResNet18 models as feature extractors to obtain deep learning features from all convolutional layers for each CT section. Since the ResNet18 network produces 3904 feature channels across all convolutional layers, each image can yield 3904 features when aggregating outputs from all layers. To obtain patient-level representations, we averaged section-level features across all tumor-containing sections and corresponding VAT sections. We then concatenated tumor and VAT features to yield fused features integrating both tissues. Feature selection involved a two-step process: We performed the Mann-Whitney U test to identify statistically significant features associated with PM. Next, we used maximum relevance and minimum redundancy to reduce redundancy and retain the top 100 nonredundant features, representing the most discriminative subset of the high-dimensional feature space. We used a sparse Bayesian extreme learning machine (SBELM) for classification. We constructed three deep learning signatures (DLS): the tumor tissue-based (TT)-DLS, VAT-based (VAT)-DLS, and fused tissue-based (F)-DLS. All feature selection and model training were performed exclusively in the training set, and the finalized model was directly applied to the internal and external test sets without retraining. Clinical Model and Multimodal Deep Learning Radiomics Model Construction We conducted univariable analysis to identify statistically significant clinical variables associated with PM. We input these variables into the SBELM model to construct a clinical model (CM). We built a multimodal deep learning radiomics model (MDLR) by integrating the F-DLS with preoperative clinical indicators using an integrated SBELM (ISBELM) ( Fig 2B – 2D ). Classifier Comparison and Radiomics Baseline Construction To justify the choice of the classifier used in the MDLR framework, we compared the ISBELM with several commonly used classifiers, including multilayer perceptron, least absolute shrinkage and selection operator, support vector machine, random forest, extreme gradient boosting, κ-nearest neighbor, and ElasticNet. All classifiers were trained and evaluated using the same training, internal test, and external test sets for a fair comparison. In addition, a traditional radiomics baseline model was constructed for comparison. Handcrafted radiomics features were extracted using the PyRadiomics package from the original, Laplacian of Gaussian-filtered, and wavelet-transformed images. Feature selection was performed sequentially using the Mann-Whitney U test, Spearman correlation filtering to remove redundant features, and regression of the least absolute shrinkage and selection operator. The selected radiomics features were then used to train an ISBELM classifier. Detailed descriptions of the classifier design, comparison experiments, and radiomics baseline construction, along with complete results, are provided in Appendix S4 . Performance Evaluation We evaluated model performance through receiver operating characteristic (ROC) curve analysis and area under the ROC curve (AUC) value calculation, with corresponding 95% CIs. We performed decision curve analysis to quantify clinical net benefit across thresholds. We used calibration curves and the Hosmer-Lemeshow test to assess model fit. We used the integrated discrimination improvement (IDI) metric to compare models. Sensitivity, specificity, and accuracy were calculated from the confusion matrix using standard definitions: sensitivity as the true positives (TP) divided by TP plus false negatives (FN); specificity as the true negatives (TN) divided by TN plus false positives (FP); and accuracy as the TP plus TN divided by TP plus TN plus FP plus FN. Visualization For exploratory visualization of section-level feature learning, we applied gradient-weighted class activation mapping (Grad-CAM) to the fine-tuned ResNet18 models of tumor and VAT CT images. Specifically, gradients of the output neuron corresponding to the predicted probability of PM were calculated with respect to the final convolutional block (layer four). Channel-wise gradients were globally averaged to generate weights, which were then linearly combined with the feature maps. The weighted maps were passed through rectified linear unit activation, bilinearly upsampled to the original section resolution, and normalized to the range [0,1]. These heatmaps provide qualitative insights into which image regions most strongly influenced section-level feature representations; however, the insights should not be interpreted as direct explanations of the final patient-level classifier. Statistical Analysis Data are presented as the means ± SDs or number (percentage). We compared clinical and pathologic variables across the training, internal test, and external test sets, as well as between patients with and without PM in the training set. We used independent Student t tests to compare continuous variables. We analyzed categorical variables with the Pearson χ 2 test, Fisher exact test, or Yates correction as appropriate. We considered a two-tailed P value less than .05 as statistically significant. We stratified patients into high- and low-risk groups based on their predicted probabilities from the MDLR. We determined the optimal cutoff value in the training set by using the Youden index from the ROC curve, and then we applied this cutoff to the internal and external test sets for risk group classification. We conducted Kaplan-Meier survival analysis to assess prognostic stratification and compared differences in RFS between risk groups using the log-rank test. The sample size was determined by the number of eligible patients available during the study period. X.W. conducted all statistical analyses with R version 4.4.1 (R Foundation for Statistical Computing), MATLAB 2020b (The MathWorks), and IBM SPSS Statistics 20.0 (SPSS). Results Patient Clinical Characteristics Among the 553 patients with pathologically confirmed GC with serosal invasion who underwent radical gastrectomy, 137 were excluded. We included a total of 416 patients (mean age, 56.6 years ± 11.6; 275 male, 141 female patients) from two centers in this study. The mean follow-up time at center 1 was 45.1 months ± 39.0, with a PM rate of 23.5% (74 of 302). The mean follow-up at center 2 was 39 months ± 27.3, with a PM rate of 22.8% (26 of 114). We divided patients into a training set ( n = 181, center 1), internal test set ( n = 121, center 1), and external test set ( n = 114, center 2). Among the 100 patients who developed PM, PM occurred either before or concurrently with recurrence at other sites. Clinicopathology variables were compared across the training, internal test, and external test sets ( Table 1 ), as well as between patients with and without PM in the training set ( Table 2 ). Among these, there were statistically significant differences in sex and tumor location among the PM and non-PM groups. The PM group included a lower proportion of male patients compared with the non-PM group (57.8% vs 75.0%, P = .03). Tumor location also differed statistically significantly between groups ( P = .01), with a higher proportion of cardia tumors (15.6% vs 6.6%) and tumors involving two or more sites (6.7% vs 0.8%), in addition to a lower proportion of antral tumors (48.9% vs 71.3%) in the PM group. Therefore, we incorporated these variables into the SBELM model to construct the CM. Table 1: Clinicopathology Characteristics of Patients in the Training and Test Sets Characteristic Total ( n = 416) Training Set ( n = 181) Test Set P Value Internal ( n = 121) External ( n = 114) Age (y) 56.6 ± 11.6 55.0 ± 11.2 55.4 ± 11.0 60.3 ± 12.1 <.001 Age range (y) 21–84 26–81 21–83 27–84 Sex Male 275 (66.1) 128 (70.7) 76 (62.8) 71 (62.3) .22 Female 141 (33.9) 53 (29.3) 45 (37.2) 43 (37.7) CEA level Normal 313 (75.2) 134 (74.0) 84 (69.4) 95 (83.3) <.001 Elevated 44 (10.6) 12 (6.6) 13 (10.8) 19 (16.7) Unknown 59 (14.2) 35 (19.4) 24 (19.8) 0 (0) Tumor location Cardia 38 (9.1) 16 (8.9) 10 (82.6) 12 (10.5) <.001 Body 91 (21.9) 42 (23.2) 33 (27.3) 16 (14.0) Antrum 264 (63.5) 119 (65.8) 73 (60.3) 72 (63.2) ≥2 sites 23 (5.5) 4 (22.1) 5 (6.8) 14 (12.3) Differentiation Poor 257 (61.8) 162 (89.5) 109 (90.1) 86 (75.4) <.001 Moderate or well 59 (14.2) 19 (10.5) 12 (9.9) 28 (24.6) N stage N0 98 (23.6) 44 (24.3) 35 (28.9) 19 (16.6) <.001 N1 73 (17.5) 41 (22.7) 22 (18.2) 10 (8.8) N2 97 (23.3) 42 (23.2) 31 (25.6) 24 (21.0) N3a 102 (24.5) 39 (21.5) 20 (16.5) 43 (37.8) N3b 46 (11.1) 15 (8.3) 13 (10.8) 18 (15.8) TNM stage II 95 (22.8) 44 (24.3) 35 (28.9) 16 (14.0) .02 III 321 (77.2) 137 (75.7) 86 (71.1) 98 (86.0) Positive lymph node ratio 0.22 (0.04–0.49) 0.19 (0.04–0.44) 0.17 (0.00–0.36) 0.34 (0.10–0.69) <.001 Record of postoperative chemotherapy No 123 (29.6) 41 (22.7) 25 (20.7) 57 (50) <.001 Yes 293 (70.4) 140 (77.3) 96 (79.3) 57 (50) Peritoneal metastasis Yes 100 45 29 26 No 316 136 92 88 Open in a new tab Note.—Unless otherwise indicated, data are presented as numbers of patients or means ± SDs, with percentages in parentheses. P values were calculated by comparing the training, internal test, and external test sets with independent Student t tests for continuous variables and the Pearson χ 2 test, Fisher exact test, or Yates correction for categorical variables. P < .05 was considered statistically significant. Table 2: Clinicopathology Characteristics of Patients with and without PM in the Training Set Characteristic Training Set ( n = 181) P Value Non-PM ( n = 136) PM ( n = 45) Age (y) 55.65 (11.3) 53.18 (10.8) .19 Sex Male 102 (75) 26 (57.8) .03 Female 34 (25) 19 (42.2) CEA level Normal 98 (72.1) 36 (80) .43 Elevated 11 (8.1) 1 (2.2) Unknown 27 (19.8) 8 (17.8) Tumor location Cardia 9 (6.6) 7 (15.6) .01 Body 29 (21.3) 13 (28.8) Antrum 97 (71.3) 22 (48.9) ≥2 sites 1 (0.8) 3 (6.7) Differentiation Poor 120 (25.7) 42 (93.3) .49 Moderate or well 16 (11.8) 3 (6.7) N stage N0 35 (25.7) 9 (20) .51 N1 32 (23.5) 9 (20) N2 33 (24.3) 9 (20) N3a 25 (18.4) 14 (31.1) N3b 11 (8.1) 4 (8.9) TNM stage II 35 (25.7) 9 (20) .42 III 101 (74.3) 36 (80) Positive lymph node ratio 0.25 (0.3) 0.34 (0.3) .046 Record of postoperative chemotherapy No 34 (25) 7 (15.6) .27 Yes 102 (75) 38 (84.4) Open in a new tab Note.—Data are presented as means ± SDs (for age and lymph node ratio) or number of patients, with percentages in parentheses. P values were calculated by comparing patients with and without peritoneal metastasis (PM) in the training set with independent Student t tests for continuous variables and the Pearson χ 2 test, Fisher exact test, or Yates correction for categorical variables. P < .05 was considered statistically significant. CEA = carcinoembryonic antigen. Performance of DLS The ROC curves and predictive performance of the deep learning models are shown in Figure 3 and Table 3 , respectively. The F-DLS achieved AUC values of 0.81 (95% CI: 0.73, 0.88) in the internal test set, 0.79 (95% CI: 0.71, 0.86) in the external test set, and 0.80 (95% CI: 0.75, 0.85) in the overall test set. Compared with the TT-DLS and VAT-DLS, the F-DLS showed a numerically higher AUC, although the differences were not statistically significant based on the DeLong test ( P > .05). In the overall test set, IDI values demonstrated that the F-DLS outperformed the TT-DLS (IDI = 0.117; 95% CI: 0.037, 0.198; P < .05) and VAT-DLS (IDI = 0.134; 95% CI: 0.069, 0.200; P < .001). Figure 3: Open in a new tab Receiver operating characteristic (ROC) curves for the deep learning models, clinical model (CM), and multimodal deep learning radiomics model (MDLR). ROC curves of the deep learning models, CM, and MDLR in the (A) training set, (B) internal test set, (C) external test set, and (D) overall test set. AUC = area under the ROC curve, F-DLS = fused tumor and VAT deep learning signature, TT-DLS = tumor tissue–based DLS; VAT-DLS = visceral adipose tissue–based DLS. Table 3: Diagnostic Performance of Deep Learning Models across Different Sets Dataset and Model AUC Sensitivity Specificity Accuracy Training set ( n = 181) TT-DLS 0.77 (0.70, 0.83) 78 (35/45) 66 (90/136) 69 (125/181) VAT-DLS 0.74 (0.67, 0.80) 78 (35/45) 65 (88/136) 68 (123/181) F-DLS 0.85 (0.79, 0.90) 73 (33/45) 83 (113/136) 81 (146/181) CM 0.60 (0.53, 0.67) 51 (23/45) 71 (96/136) 66 (119/181) MDLR 0.87 (0.82, 0.92) 80 (36/45) 82 (112/136) 82 (148/181) Internal test set ( n = 121) TT-DLS 0.71 (0.63, 0.79) 83 (24/29) 58 (53/92) 64 (77/121) VAT-DLS 0.69 (0.60, 0.77) 72 (21/29) 67 (62/92) 69 (83/121) F-DLS 0.81 (0.73, 0.88) 69 (20/29) 77 (71/92) 75 (91/121) CM 0.57 (0.47, 0.66) 45 (13/29) 56 (52/92) 54 (65/121) MDLR 0.86 (0.79, 0.92) 86 (25/29) 80 (74/92) 82 (99/121) External test set ( n = 114) TT-DLS 0.72 (0.63, 0.80) 81 (21/26) 46 (40/88) 54 (61/114) VAT-DLS 0.73 (0.64, 0.81) 77 (20/26) 56 (49/88) 61 (69/114) F-DLS 0.79 (0.71, 0.86) 62 (16/26) 80 (70/88) 75 (86/114) CM 0.55 (0.45, 0.64) 46 (12/26) 59 (52/88) 56 (64/114) MDLR 0.86 (0.78, 0.92) 77 (20/26) 80 (70/88) 79 (90/114) Overall test set ( n = 235) TT-DLS 0.71 (0.65, 0.77) 82 (45/55) 52 (93/180) 59 (138/235) VAT-DLS 0.70 (0.64, 0.76) 75 (41/55) 62 (111/180) 65 (152/235) F-DLS 0.80 (0.75, 0.85) 66 (36/55) 78 (141/180) 75 (177/235) CM 0.55 (0.49, 0.62) 46 (25/55) 58 (104/180) 55 (129/235) MDLR 0.86 (0.80, 0.90) 82 (45/55) 80 (144/180) 80 (189/235) Open in a new tab Note.—Values represent area under the receiver operating characteristic curve (AUC, with 95% CIs in parentheses), sensitivity, specificity, and accuracy calculated within each dataset (training, internal test, external test, and overall test). Integrated discrimination improvement values for the comparison of the fused tumor deep learning signature (F-DLS) and visceral adipose tissue (VAT) DLS against tumor tissue—based (TT)-DLS and VAT-DLS in the overall test set are also provided. CM = clinical model, MDLR = multimodal deep learning radiomics model. Classifier Comparison and Radiomics Baseline Performance As summarized in Table S1 and illustrated in Figure S1 , ISBELM consistently achieved superior predictive performance and generalization across the training, internal test, and external test sets. In addition, compared with the radiomics signature, the proposed deep learning–based models, particularly the MDLR, demonstrated higher AUC values across all test sets, indicating improved predictive capability over conventional radiomics approaches ( Table S1 ). Comparison with Clinical and Multimodal Models The F-DLS outperformed the CM across all datasets (AUC range, 0.68–0.74). In the overall test set, the IDI for the comparison between the F-DLS and the CM was 0.139 (95% CI: 0.065, 0.213; P < .001). We developed an MDLR integrating the F-DLS with clinical indicators (sex, tumor location) using the ISBELM model. The MDLR achieved AUCs of 0.86 (95% CI: 0.79, 0.92), 0.86 (95% CI: 0.78, 0.92), and 0.86 (95% CI: 0.80, 0.90) in the internal, external, and overall test sets, respectively. In the overall test set, the MDLR outperformed both the CM (AUC, 0.86 vs 0.55; IDI, 0.279; 95% CI: 0.22, 0.33; P < .001) and the F-DLS (AUC, 0.86 vs 0.80; IDI, 0.066; 95% CI: 0.02, 0.11; P = .002). Analysis of calibration curves revealed strong agreement between predicted and observed outcomes ( Fig S2 ). The Hosmer-Lemeshow test indicated good model fit ( P > .05). Decision curve analysis revealed that the MDLR provided the greatest net benefit, especially in the threshold range of 0.37–0.60 ( Fig S3 ). Prognostic Stratification Based on the MDLR Analysis of Kaplan-Meier survival curves revealed that the MDLR successfully stratified patients into high- and low-risk groups for PM ( Fig 4 ). We observed statistically significant differences in RFS across all three sets ( P ≤ .001 for all comparisons). In the overall set, high-risk patients identified by the MDLR had statistically significant shorter RFS than low-risk patients (mean RFS, 30.0 months ± 27.3 vs 46.1 months ± 32.0). Figure 4: Open in a new tab Kaplan-Meier survival analysis based on the multimodal deep learning radiomics model (MDLR). Kaplan-Meier curves for recurrence-free survival in the (A) training set, (B) internal test set, (C) external test set, and (D) overall test set, stratified according to the MDLR-predicted risk groups. Patients classified as high risk by the MDLR have statistically significantly shorter recurrence-free survival than those in the low-risk group across all sets (log-rank P ≤ .001 for all comparisons). Interpretability Analysis Based on Grad-CAM To further interpret the model’s decision-making process, we generated Grad-CAM heatmaps based on the preoperative CT images. In patients who subsequently developed PM, analysis of activation maps revealed that the model predominantly focused on the tumor-serosa interface and adjacent heterogeneous VAT regions. In contrast, patients without PM exhibited more diffuse or less intense activation patterns without distinct focal areas ( Fig S4 ). Discussion Accurately assessing the risk of PM after radical gastrectomy is critical for guiding treatment strategies and improving outcomes in patients with GC with serosal invasion. In this study, we developed a MDLR based on preoperative CT images and clinical indicators to predict PM risk. The proposed MDLR model demonstrated robust predictive performance, with AUCs of 0.86 (95% CI: 0.79, 0.92) and 0.86 (95% CI: 0.78, 0.92) in the internal and external test sets, respectively. In the overall test set, it statistically significantly outperformed the CM (IDI, 0.279; 95% CI: 0.224, 0.334; P < .001) and the F-DLS (IDI, 0.066; 95% CI: 0.024, 0.109; P = .002). These findings suggest that the proposed model may serve as a useful tool for early identification of patients with GC with serosal invasion who are at high risk of PM. PM in GC remains a major clinical challenge owing to its high incidence, poor prognosis, and lack of reliable early diagnostic markers. PM occurs through a complex multistep process involving tumor serosal invasion and tumor cell detachment, peritoneal dissemination, and subsequent implantation on peritoneal surfaces. These biologic behaviors are often silent in the early stage, making timely detection particularly difficult. Despite curative resection, micrometastatic disease may already be present at the time of surgery, underscoring the need for effective risk stratification tools. Several models based on radiomics and deep learning have been proposed for preoperative detection of occult PM, with promising accuracy ( 20 , 21 ). However, most of these studies focused on preoperative diagnosis of radiologically undetectable PM, rather than predicting PM onset following radical surgery in patients already known to have serosal invasion. Currently, PM diagnosis after surgery relies on imaging and clinical symptoms, which often reflect advanced disease. In a recent study, a competing-risk nomogram incorporating a collagen signature and pathology features to predict PM was proposed ( 3 ), but the nomogram is dependent on multiphoton imaging, which is a technically demanding and not widely available imaging modality. In contrast, CT imaging is a standard, noninvasive, and broadly applicable imaging modality. In this study, we propose a CT-based deep learning model in which both tumor and VAT features are used for PM risk prediction. Although our end point was postoperative PM, the model was based on preoperative CT and clinical indicators and can therefore be used before surgery. While higher TNM staging and lymph node status are generally associated with increased metastatic risk, their discriminative value for predicting PM may be limited in our cohort, in which all patients with GC included in our study had serosal invasion (T4a), resulting in a relatively homogeneous high-risk population and reducing the discriminatory power of stage-related variables for predicting PM. Notably, our model enabled risk stratification within this homogeneous T4a group, identifying those most likely to develop PM. Beyond risk prediction, this approach could support clinical decision-making by identifying patients of high risk who may benefit from intraoperative IPC or extended peritoneal lavage cytologic analysis. This approach may also be used to guide postoperative strategies, such as intensified adjuvant therapy, clinical trial enrollment, or closer surveillance, and facilitate informed discussions with patients regarding treatment expectations and disease prognosis. Accumulating evidence suggests that VAT is not merely a passive fat depot but also plays an active role in tumor progression through immunologic and metabolic mechanisms ( 22 ). VAT harbors various immune cells and secreted cytokines that modulate the tumor microenvironment. Recent studies have revealed that peritoneal adipocytes, particularly omental adipocytes, can acquire cancer-associated fibroblast-like properties and promote tumor progression. For example, Natsume et al ( 23 ) demonstrated that omental adipocyte-derived CXC motif chemokine 2 activates the RAC-gamma serine and threonine-protein kinase signaling pathway and increases vascular endothelial growth factor-A expression in GC cells, while Hamabe-Horiike et al reported that omental adipocytes exposed to GC cells undergo dedifferentiation and secrete protumorigenic factors ( 14 ). These findings suggest that VAT microenvironmental alterations may facilitate peritoneal implantation and survival of disseminated tumor cells. Li et al observed statistically significant differences in VAT attenuation between patients who are PM-positive and PM-negative and correlations with tumor stage and serum markers ( 24 ). However, they did not observe that VAT metrics alone were independent predictors of PM, possibly because of oversimplified measurements. In our study, we refined this approach by segmenting VAT at tumor-matched sections and extracting high-level features via deep learning. Our VAT-DLS model showed good predictive value, while the combined tumor-VAT model (F-DLS) outperformed both single-region models. The improved performance of the F-DLS may reflect synergistic interactions between tumor characteristics and the surrounding VAT microenvironment. Such interactions may involve stromal remodeling, inflammatory responses, or metabolic cross-talk at the tumor-VAT interface, which are difficult to capture with routine visual assessment on CT images but may be recognized by deep learning. This observation supports the biologic plausibility of our approach and underscores the value of integrating both tumor and VAT features in risk prediction. To the best of our knowledge, ours is the first study to use deep learning for modeling the relationship between VAT and PM after radical gastrectomy. Although deep learning methods remain somewhat of a black box, they provide a powerful way to extract latent patterns that go beyond what routine, visually assessed radiologic features can capture. Instead of relying solely on CT attenuation values, which are sometimes insufficiently sensitive, the ResNet18-based network captures multiscale image information that might reflect microscopic immune or structural changes in the VAT and tumor microenvironment. However, the proposed deep learning framework currently relies on manual tumor segmentation, which requires substantial human effort and may limit scalability. Incorporating automated segmentation techniques in future work may help streamline the process and increase feasibility of use in routine practice. To increase model interpretability, we applied Grad-CAM to visualize attention regions on preoperative CT images. Analysis of the activation maps revealed that the model frequently focused not only on the tumor itself but also on adjacent peritumoral tissues and VAT zones, which is consistent with the recognized role of the tumor-adipose tissue interface and the local microenvironment in peritoneal dissemination. Distinct Grad-CAM activation patterns were observed between patients who subsequently developed PM and those who did not. In patients with PM, model attention tended to cluster around the tumorserosa interface and heterogeneous VAT regions, potentially reflecting subclinical stromal remodeling, subtle inflammatory changes, or increased microenvironmental heterogeneity. In contrast, patients without PM exhibited more diffuse, low-intensity activation patterns without well-defined focal areas, suggesting fewer early microenvironmental alterations. Although the highlighted regions did not precisely coincide with manually delineated tumor regions of interest, this is consistent with the class-discriminative rather than localization-specific nature of Grad-CAM. Importantly, the model maintained consistent predictive performance across test sets, suggesting that it captures imaging biomarkers indicative of metastatic risk that extend beyond anatomically defined tumor boundaries. Despite promising results, our study has limitations. First, our study was retrospective in nature, which may introduce bias. Second, although we included patients from two centers, the sample size was moderate because of strict inclusion criteria. Despite multistep feature selection and independent internal and external test sets, potential overfitting and limited statistical stability cannot be entirely excluded, and large prospective multicenter cohorts are warranted to validate the generalizability of our findings. Third, we limited inclusion to patients with pathologically proven serosa-invasive GC to ensure a homogeneous high-risk cohort. While this strengthened internal validity, it may have reduced the applicability of our findings to the broader GC population. Nevertheless, because the model was trained on preoperative CT data, it can be applied prospectively once serosal invasion is depicted radiologically, and the integration of serosal invasion assessment and PM prediction into a unified framework should be an aim of future work. Finally, while our model provides predictive insights, the biologic relevance of the extracted features remains to be further elucidated—particularly how VAT characteristics mechanistically influence PM development in GC. In addition, although Grad-CAM was used to provide qualitative visualization of section-level feature learning, these heatmaps should not be interpreted as direct explanations of the final patient-level classifier, given the aggregation of section-level features and subsequent modeling steps. In conclusion, we propose a novel deep learning model incorporating VAT and tumor features from preoperative CT images to predict PM risk in patients with GC with serosal invasion. The model holds promise for personalized decision-making, potentially guiding the selection of candidates for IPC or closer postoperative monitoring. Supplemental Files Appendices S1-S4, Tables S1-S2, Figures S1-S7 rycan250353suppa1.pdf (1.2MB, pdf) Conflicts of Interest rycan250353coi.zip (529.5KB, zip) * Y.L. and X.W. contributed equally to this work. Funding: This work was supported by the National Key Research and Development Program of China (grant no. 2023YFF1204600), Medical Scientific Research Foundation of Guangdong Province, China (grant no. B2025305), the National Natural Science Foundation of China (grant no. 82402262), the Fundamental Research Funds for the Central Universities (grant no. 21624304), and the Guangzhou Science and Technology Program (grant no. 2025A04J3480). Abbreviations: AUC area under the ROC curve CM clinical model DLS deep learning signature F-DLS fused tumor-VAT DLS GC gastric cancer Grad-CAM gradient-weighted class activation mapping IDI integrated discrimination improvement IPC intraperitoneal chemotherapy ISBELM integrated SBELM MDLR multimodal deep learning radiomics model PM peritoneal metastasis RFS recurrence-free survival ROC receiver operating characteristic SBELM sparse Bayesian extreme learning machine TT-DLS tumor tissue–based DLS VAT visceral adipose tissue VAT-DLS visceral adipose tissue–based DLS Disclosures of conflicts of interest Please see ICMJE form(s) for author conflicts of interest. These have been provided as supplemental materials . References 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 ( 6 ): 394 – 424 . [ DOI ] [ PubMed ] [ Google Scholar ] 2. Ikoma N , Chen HC , Wang X , et al . Patterns of initial recurrence in gastric adenocarcinoma in the era of preoperative therapy . Ann Surg Oncol 2017. ; 24 ( 9 ): 2679 – 2687 . [ DOI ] [ PubMed ] [ Google Scholar ] 3. Chen D , Liu Z , Liu W , et al . Predicting postoperative peritoneal metastasis in gastric cancer with serosal invasion using a collagen nomogram . Nat Commun 2021. ; 12 ( 1 ): 179 . [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 4. Wang FH , Zhang XT , Tang L , et al . The Chinese Society of Clinical Oncology (CSCO): clinical guidelines for the diagnosis and treatment of gastric cancer, 2023 . Cancer Commun (Lond) 2024. ; 44 ( 1 ): 127 – 172 . [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Thomassen I , van Gestel YR , van Ramshorst B , et al . Peritoneal carcinomatosis of gastric origin: a population-based study on incidence, survival and risk factors . Int J Cancer 2014. ; 134 ( 3 ): 622 – 628 . [ DOI ] [ PubMed ] [ Google Scholar ] 6. Van Cutsem E , Sagaert X , Topal B , Haustermans K , Prenen H

. Gastric cancer . Lancet 2016. ; 388 ( 10060 ): 2654 – 2664 . [ DOI ] [ PubMed ] [ Google Scholar ] 7. Gastric Cancer Committee of China Anti Cancer Association . Chinese expert consensus on the diagnosis and treatment of peritoneal metastasis of gastric cancer (2023 edition) . Zhonghua Wei Chang Wai Ke Za Zhi 2023. ; 26 ( 8 ): 717 – 728 . [ DOI ] [ PubMed ] [ Google Scholar ] 8. Yang XJ , Huang CQ , Suo T , et al . Cytoreductive surgery and hyperthermic intraperitoneal chemotherapy improves survival of patients with peritoneal carcinomatosis from gastric cancer: final results of a phase III randomized clinical trial . Ann Surg Oncol 2011. ; 18 ( 6 ): 1575 – 1581 . [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Mi DH , Li Z , Yang KH , et al . Surgery combined with intraoperative hyperthermic intraperitoneal chemotherapy for gastric cancer: a systematic review and meta-analysis of randomised controlled trials . Int J Hyperthermia 2013. ; 29 ( 2 ): 156 – 167 . [ DOI ] [ PubMed ] [ Google Scholar ] 10. Glehen O , Kwiatkowski F , Sugarbaker PH , et al . Cytoreductive surgery combined with perioperative intraperitoneal chemotherapy for the management of peritoneal carcinomatosis from colorectal cancer: a multi-institutional study . J Clin Oncol 2004. ; 22 ( 16 ): 3284 – 3292 . [ DOI ] [ PubMed ] [ Google Scholar ] 11. 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 ( 5 ): 808 – 816 . [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 12. Sun F , Feng M , Guan W

. Mechanisms of peritoneal dissemination in gastric cancer . Oncol Lett 2017. ; 14 ( 6 ): 6991 – 6998 . [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Yonemura Y , Ishibashi H , Mizumoto A , et al . The development of peritoneal metastasis from gastric cancer and rationale of treatment according to the mechanism . J Clin Med 2022. ; 11 ( 2 ): 458 . [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Hamabe-Horiike T , Harada SI , Yoshida K , Kinoshita J , Yamaguchi T , Fushida S

. Adipocytes contribute to tumor progression and invasion of peritoneal metastasis by interacting with gastric cancer cells as cancer-associated fibroblasts . Cancer Rep (Hoboken) 2023. ; 6 ( 1 ): e1647 . [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 15. Huang H , Yang X , Sun J , et al . Value of visceral fat area in the preoperative discrimination of peritoneal metastasis from gastric cancer in patients with different body mass index: a prospective study . Cancer Manag Res 2020. ; 12 : 6523 – 6532 . [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 16. Catanese S , Aringhieri G , Vivaldi C , et al . Role of baseline computed-tomography-evaluated body composition in predicting outcome and toxicity from first-line therapy in advanced gastric cancer patients . J Clin Med 2021. ; 10 ( 5 ): 1079 . [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Lee JJ , Pedley A , Hoffmann U , et al . Cross-sectional associations of computed tomography (CT)-derived adipose tissue density and adipokines: the Framingham Heart Study . J Am Heart Assoc 2016. ; 5 ( 3 ): e002545 . [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 18. Bian L , Wu D , Chen Y , et al . Associations of radiological features of adipose tissues with postoperative complications and overall survival of gastric cancer patients . Eur Radiol 2022. ; 32 ( 12 ): 8569 – 8578 . [ DOI ] [ PubMed ] [ Google Scholar ] 19. Jiang Y , Jin C , Yu H , et al . Development and validation of a deep learning CT signature to predict survival and chemotherapy benefit in gastric cancer: a multicenter, retrospective study . Ann Surg 2021. ; 274 ( 6 ): e.1153 – e1161 . [ DOI ] [ PubMed ] [ Google Scholar ] 20. Jiang Y , Liang X , Wang W , et al . Noninvasive prediction of occult peritoneal metastasis in gastric cancer using deep learning . JAMA Netw Open 2021. ; 4 ( 1 ): e2032269 . [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 21. Dong D , Tang L , Li ZY , et al . Development and validation of an individualized nomogram to identify occult peritoneal metastasis in patients with advanced gastric cancer . Ann Oncol 2019. ; 30 ( 3 ): 431 – 438 . [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 22. Lee JW , Son MW , Chung IK , Cho YS , Lee MS , Lee SM

. Significance of CT attenuation and F-18 fluorodeoxyglucose uptake of visceral adipose tissue for predicting survival in gastric cancer patients after curative surgical resection . Gastric Cancer 2020. ; 23 ( 2 ): 273 – 284 . [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 23. Natsume M , Shimura T , Iwasaki H , et al . Omental adipocytes promote peritoneal metastasis of gastric cancer through the CXCL2-VEGFA axis . Br J Cancer 2020. ; 123 ( 3 ): 459 – 470 . [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 24. Li LM , Feng LY , Liu CC , et al . Can visceral fat parameters based on computed tomography be used to predict occult peritoneal metastasis in gastric cancer? World J Gastroenterol 2023. ; 29 ( 15 ): 2310 – 2321 . [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Appendices S1-S4, Tables S1-S2, Figures S1-S7 rycan250353suppa1.pdf (1.2MB, pdf) Conflicts of Interest rycan250353coi.zip (529.5KB, zip) Articles from Radiology: Imaging Cancer are provided here courtesy of Radiological Society of North America ACTIONS View on publisher site PDF (1.8 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• CT semantic features of primary gastric cancer: a preoperative predictive model for peritoneal metastasis with superior efficacy to conventional direct CT assessment.#251089
• Deep-learning endomicroscope with large field-of-view and depth-of-field for real-time in vivo imaging of epithelial cancer hallmarks.#214008
• Prospective Validation of an AI Model for Predicting Liver Metastasis in Colorectal Cancer#857883
• Novel Humanized Ferritin-based NIR Fluorescent Molecular Probe for Identifying Tumor Margins in Gastric Tissue#900326
• GutCore: An Endoscopy Foundation Model for Whole-Case Gastric Cancer Analysis#459880