Predicting Prognosis for Gastric Cancer Patients Receiving Neoadjuvant Treatment With Body Composition‐Based Deep Learning

ABSTRACT

Background

This study sought to develop an innovative body composition (BC)‐based deep learning (DL) model to precisely evaluate survival in gastric cancer (GC) patients undergoing neoadjuvant treatment (NT).

Materials and Methods

This retrospective study included GC patients undergoing NT from two centers. CT images both pre‐NT and post‐NT were preprocessed, focusing on the automatic segmentation of subcutaneous fat, visceral fat, and skeletal muscle regions using TotalSegmentator. Delta Radiomics features were extracted using Pyradiomics. After feature fusion and selection, the optimal model is Naive Bayes (Rad model). A hybrid DL model was developed by combining ResNet18 and Transformer networks for feature extraction. The Clinic_Rad_DL model was constructed by combining clinical features, radiomic signatures, and DL signatures. The ExtraTree classifier was used for the Clinic_Rad_DL model, while a separate Cox regression model was developed for survival analysis using the same features.

Results

A total of 356 patients (mean age, 59 ± 10 years; 264 males [74.2%]) were enrolled and divided into training, validation, and test sets in a 7:2:1 ratio. The DL model outperformed the Rad model. The Clinic_Rad_DL model outperformed both Rad model and DL model, with AUC of 0.915, 0.890, and 0.890 in training, validation, and test sets, respectively. The Cox proportional hazards model showed C‐index of 0.806, 0.803, and 0.819, effectively stratifying patients into high‐ and low‐risk groups with significant survival differences.

Conclusion

The study developed and validated a BC‐based DL model to predict survival in GC patients undergoing NT, offering potential for personalized treatment strategies in clinical practice.

1. Introduction

Gastric cancer (GC) is the fifth most common malignancy and the fifth leading cause of cancer‐related death worldwide. In China, GC remains a major health burden, with over 358,000 new cases and 260,000 deaths annually, accounting for approximately 40% of GC‐related deaths worldwide [1, 2]. Around 80% of cases are diagnosed at an advanced stage, resulting in a poor prognosis [3]. Improving outcomes for these patients is imperative, albeit challenging. Neoadjuvant treatment (NT) can downstage tumors and increase the rate of R0 resection, thus improving survival for patients with locally advanced GC [4, 5].

In addition to traditional data used for outcome prediction in GC patients, body composition (BC) analysis is garnering heightened interest. Our previous research demonstrated that skeletal muscle and adipose tissue metrics from computed tomography (CT) scans are associated with various outcomes in GC patients undergoing radical gastrectomy following NT, with potential clinical implications [6, 7, 8]. Nonetheless, this approach is limited by its reliance on BC measurements from a single CT slice at a specific anatomical site (the third lumbar vertebra level, L3 level). Therefore, a more comprehensive BC evaluation may be warranted.

CT plays a vital role in the diagnosis and staging of GC as well as assess treatment efficacy. Furthermore, CT offers valuable insights into tumor behavior and prognosis. Radiomics, a technique that analyzes medical images using mathematical methods, extracts Supporting Information by analyzing signal intensities in the spatial distribution and relationships between pixels. This quantitative approach facilitates a more comprehensive understanding of image features (shape, size, intensity, and texture) [9, 10]. Radiomics parameters quantitatively reflect the heterogeneity of the analyzed regions of interest and the underlying pathophysiological changes [11, 12].

Recent studies suggest that radiomic features of skeletal muscle and adipose tissue may offer greater sensitivity in predicting cancer outcomes compared to traditional BC analysis [13, 14, 15]. Developing robust radiomics‐based conclusions typically necessitates high=quality and large‐scale datasets, for which deep learning (DL) can provide substantial reinforcement. As a subset of machine learning, the neural network architecture inherent in DL allows models to expand exponentially in response to increasing data volumes and complexities [16]. Currently, DL has been extensively developed for applications in disease diagnosis, particularly alongside omics technologies, due to its ability to process large datasets.

Current prognostic stratification for GC patients predominantly relies on the TNM tumor staging system. Nonetheless, among GC patients who received NT, prognostic variations have been noted even among those with identical postoperative pathological stages. Given the increasing use of NT among GC patients, there is an imperative need to optimize treatment strategies based on individualized patient risk profiles. Therefore, this study aimed to develop and validate an innovative BC‐based DL model and a comprehensive model combining clinical, radiomics, and DL to predict OS in GC patients who received NT, utilizing both pre‐NT and post‐NT CT images.

2. Materials and Methods

2.1. Patient Population

This retrospective study included GC patients who underwent NT between May 2014 and December 2018 at Fudan University Shanghai Cancer Center (Institution I) and Peking Union Medical College Hospital (Institution II). Patients with esophagogastric junction adenocarcinoma who underwent neoadjuvant chemoradiotherapy were also included. One of the following neoadjuvant chemotherapy regimens was administered: SOX (oxaliplatin and S‐1), XELOX (oxaliplatin and capecitabine), or FOLFOX (5‐FU, oxaliplatin, and leucovorin).

The inclusion criteria were: (1) confirmed diagnosis of gastric adenocarcinoma via endoscopic biopsy; (2) completion of radical gastrectomy and D2 lymphadenectomy following NT; and (3) availability of abdominal CT examinations performed at the same hospital both before NT and after NT. The exclusion criteria were: (1) presence of distant metastasis; (2) incomplete clinicopathological or follow‐up data; (3) diagnosis of gastric stump cancer; (4) history of another neoplasia within the last 5 years; (5) presence of other severe complications. The research protocol was approved by the Institutional Review Board of Peking Union Medical College Hospital (Approval No. JS‐2655) and Fudan University Shanghai Cancer Center (Approval No. 1612167–18). All patients provided written informed consent.

A total of 356 GC patients (200 from Institution I and 156 from Institution II) were enrolled in the final analysis. The dataset was stratified through random sampling into a training set (n = 249), a validation set (n = 72), and a testing set (n = 35) at a ratio of 7:2:1. The overall workflow of this study is illustrated in Figure 1.

FIGURE 1.

Overview of the study design and workflow. (A) Detailed flowchart of the patient selection process. (B) Schematic representation of the model construction pipeline.

2.2. Clinical Data Collection and CT Scanning Parameters

Clinical, treatment, and pathological data were retrieved from the electronic medical record system of the participating institutions. All GC patients were followed up via outpatient visits or telephone interviews until death or last follow‐up. Overall survival (OS) was calculated from the date of surgical operation to the date of death (all‐cause mortality) or the date of the last follow‐up.

All patients underwent abdominal CT within the 4 weeks prior to NT initiation and within the 4 weeks prior to surgery. All CT images were acquired from the picture archiving and communication system (PACS) and saved in DICOM format for further evaluation. The CT scanners and primary scanning parameters for each institution are detailed in Table S1.

2.3. Abdominal Fat and Skeletal Muscle Segmentation

Following the acquisition of raw DICOM data, all CT images underwent format conversion, bias field correction, and normalization. After preprocessing, automated segmentation was performed on thin‐slice images (0.625–1.5 mm thickness) from contrast‐enhanced CT scans. This study focused on CT slices of the abdominal region, specifically extending from the lung bases to the upper margin of the pelvic bone, to maintain a consistent anatomical range. TotalSegmentator, an open source Python library built on the nnU‐Net architecture, was used to automatically segment subcutaneous fat, visceral fat, and skeletal muscle. Representative examples of the automated segmentation are shown in Figure 2 [17].

FIGURE 2.

Representative images illustrating the automated segmentation of subcutaneous fat, visceral fat, and skeletal muscle.

2.4. Extraction and Selection of Delta Radiomics Features

Delta radiomics features were extracted from the subcutaneous fat region, visceral fat region, and skeletal muscle region within the 3D volumes. Using the open‐source Python package Pyradiomics, features were extracted from both unfiltered (raw image) and filtered data. Delta radiomics assess the relative net changes of radiomic features in longitudinal images, providing valuable information to identify, quantify, and potentially predict changes induced by treatment [18]. Therefore, delta features were calculated as the relative change between post‐treatment and pre‐treatment features using the following equation:

$$\mathrm{\Delta F}=\frac{\left(F_{\mathit{\text{post}}}-F_{\mathit{pre}}\right)}{F_{\mathit{pre}}}$$

Here, F _pre refers to the features extracted from the pre‐NT CT images, and F _post refers to the features extracted from the postNT CT images.

Further details regarding the extraction and selection of Delta Radiomics Features were provided in the Supporting Information.

2.5. Model Development

2.5.1. Delta Radiomic Model (Rad Model)

Following feature fusion and selection, a machine learning classification model based on delta‐radiomics features was constructed using the scikit‐learn library. The probability output generated by the optimal model is referred to as the Radiomic signature (Rad‐sign).

2.5.2. Clinical Model

Univariate and multivariate logistic regression analyses were conducted to screen for clinical risk factors associated with survival outcomes.

2.5.3. Deep Learning Model

The pre‐NT and post‐NT images are processed prior to being input into the DL network. We developed a hybrid DL model termed ResHy‐Net (ResNet‐based Hybrid Network) by integrating a ResNet18 network with a Transformer network [19, 20]. Because of the single‐channel 3D image patch input, ResNet18 did not utilize ImageNet pretrained weights to avoid domain mismatch. As depicted in Figure 1, the ResNet18 network functioned as a robust feature extraction tool. The probability output generated by the DL model is referred to as the Deep Learning signature (DL‐sign).

2.5.4. Fusion Model

The selected clinical features, DL‐sign, and Rad‐sign were concatenated into a single feature vector to establish the comprehensive fusion model.

Technical details of the model development were provided in the Supporting Information.

2.6. Statistical Analysis

Continuous variables are presented as the mean ± standard deviation or median (interquartile range), while categorical variables are expressed as counts (n) and percentages (%). Statistical analyses were conducted using Python 3.8, utilizing numpy, sklearn, and lifelines as the primary third‐party libraries. The models were developed through the PixelmedAI platform.

3. Results

3.1. Baseline Characteristics of Patients

A total of 356 patients were included, and the demographic and clinical characteristics of all patients in the cohort are presented in Table S2. The clinically relevant risk factors identified through univariate and multivariate logistic regression (p < 0.05) were the ypT stage and ypN stage. The results of the univariate and multivariate logistic regression analyses are shown in Table 1.

TABLE 1.

Univariate and multivariate logistic regression analysis.

Feature name	Univariate		Multivariate
	OR (95% CI)	p	OR (95% CI)	p
Gender (Female)	0.968 (0.599, 1.565)	0.895	——	——
Age (≤ 65)	0.834 (0.532, 1.308)	0.43	——	——
BMI (> 23)	0.667 (0.437, 1.018)	0.06	——	——
NT scheme (NCRT)	0.648 (0.364, 1.153)	0.14	——	——
Hb (M ≥ 130, F ≥ 115)	0.694 (0.451, 1.068)	0.097	——	——
ALB (≥ 35)	0.63 (0.376, 1.057)	0.08	——	——
PA (≥ 200)	0.553 (0.346, 0.882)	0.013	0.040 (0.374, 1.078)	0.093
Tumor location (Middle+Lower)	0.852 (0.677, 1.072)	0.172	——	——
Tumor size (< 5)	0.395 (0.242, 0.645)	< 0.001	0.636 (0.367, 1.10)	0.106
ypT (3–4)	5.52 (3.351, 9.091)	< 0.001	2.495 (1.360, 4.577)	0.003
ypN (1–3)	5.957 (3.72, 9.529)	< 0.001	3.221 (1.883, 5.512)	< 0.001
Pathologic response (non‐pCR)	6.167 (2.355, 16.150)	< 0.001	1.823 (0.612, 5.429)	0.281

Abbreviations: ALB, albumin; BMI, body mass index; CI, confidence interval; Hb, hemoglobin; NCRT, neoadjuvant chemoradiotherapy; NT, neoadjuvant treatment; OR, odds ratio; PA, prealbumin; pCR, pathological Complete Response; ypN, post‐neoadjuvant therapy N stage; ypT, post‐neoadjuvant therapy T stage.

3.2. Delta Radiomics Feature Selection

LASSO regression analysis was used for dimensionality reduction of radiomic features. The selection of the penalty coefficient (λ = 0.026) and the feature selection process, along with the plot showing the variation of feature coefficients with λ, are demonstrated in Figure 3. Initially, a total of 4683 features were extracted; after preliminary screening using an independent samples t‐test, 1283 features were retained; in the second step, further screening resulted in 79 features; and in the third step, LASSO regression was used to ultimately select 21 features for modeling (Figure 3).

FIGURE 3.

LASSO regression‐based radiomics feature selection. An optimal λ value of 0.026 was selected. (A) Ten‐fold cross‐validation for tuning parameter (λ) selection in the LASSO model via minimum negative log loss. The dotted vertical line represents the optimal log(λ) value. (B) LASSO coefficient profiles of the radiomics features. Each curve represents the coefficient path of a single independent variable across varying λ values. (C) Bar chart illustrating feature importance. The y‐axis lists the 32 selected features, and the x‐axis represents their corresponding nonzero coefficients in the arterial phase habitat radiomics model. LASSO, least absolute shrinkage and selection operator.

3.3. Model Performance and Validation

The optimal model in Delta radiomics is the Naive Bayes model (Rad model). The accuracy, sensitivity, specificity, and area under the curve (AUC) of the training set are 0.654, 0.785, 0.556, and 0.706 (95% CI: 0.703, 0.807), respectively; the validation set accuracy, sensitivity, specificity, and AUC are 0.597, 0.484, 0.683, and 0.549 (95% CI: 0.442, 0.701), respectively; and the test set accuracy, sensitivity, specificity, and AUC are 0.657, 0.800, 0.550, and 0.653 (95% CI: 0.482, 0.792), respectively.

The performance of the DL model is as follows: the accuracy, sensitivity, specificity, and AUC of the training set are 0.803, 0.953, 0.690, and 0.891 (95% CI: 0.848, 0.930), respectively; for the validation set, they are 0.806, 0.710, 0.878, and 0.841 (95% CI: 0.735, 0.924), respectively; and for the test set, they are 0.771, 0.667, 0.850, and 0.747 (95% CI: 0.557, 0.920), respectively.

The optimal model of the Clinic_Rad combined model is the MLP model. The accuracy, sensitivity, specificity, and AUC of the training set are 0.771, 0.682, 0.838, and 0.809 (95% CI: 0.755, 0.856), respectively; for the validation set, they are 0.736, 0.742, 0.732, and 0.756 (95% CI: 0.672, 0.894), respectively; and the test set shows accuracy, sensitivity, specificity, and AUC of 0.800, 0.600, 0.950, and 0.817 (95% CI: 0.639, 0.940), respectively.

The optimal model of the Clinic_Rad_DL combined model is the ExtraTree model. The diagnostic performances of all tested models were shown in Table S3. The accuracy, sensitivity, specificity, and AUC of the training set are 0.871, 0.785, 0.937, and 0.915 (95% CI: 0.886, 0.943), respectively; for the validation set, they are 0.833, 0.839, 0.829, and 0.890 (95% CI: 0.811, 0.953), respectively; and the test set shows accuracy, sensitivity, specificity, and AUC of 0.914, 0.933, 0.900, and 0.890 (95% CI: 0.771, 0.991), respectively (Table 2).

TABLE 2.

Comparison of model performance.

Model name	Accuracy	AUC (95% CI)	Sensitivity	Specificity	Task
Rad	0.654	0.706 (0.703, 0.807)	0.785	0.556	Train
Rad	0.597	0.549 (0.442, 0.701)	0.484	0.683	Val
Rad	0.657	0.653 (0.482, 0.792)	0.8	0.55	Test
DL(Vit)	0.803	0.891 (0.848, 0.930)	0.953	0.69	Train
DL(Vit)	0.806	0.841 (0.735, 0.924)	0.71	0.878	Val
DL(Vit)	0.771	0.747 (0.557, 0.920)	0.667	0.85	Test
Clinic_Rad	0.771	0.809 (0.755, 0.856)	0.682	0.838	Train
Clinic_Rad	0.736	0.756 (0.672, 0.894)	0.742	0.732	Val
Clinic_Rad	0.8	0.817 (0.639, 0.940)	0.6	0.95	Test
Clinic_Rad_DL	0.871	0.915 (0.886, 0.943)	0.785	0.937	Train
Clinic_Rad_DL	0.833	0.890 (0.811, 0.953)	0.839	0.829	Val
Clinic_Rad_DL	0.914	0.890 (0.771, 0.991)	0.933	0.9	Test

Among all models, the optimal model is the Clinic_Rad_DL model, which has a higher AUC in the test set compared to other models (Figure 4). The differences in AUC between the Clinic_Rad_DL model and other models in the test set are not statistically significant (Table S4). The calibration curves indicate that each model has good consistency between predicted results and actual outcomes in both the training and test sets (Figure S1), and the Hosmer–Lemeshow test p values are all greater than 0.05 (Table S5). Additionally, the decision curve analysis (DCA) curves analyzed in the training, validation, and test sets show that the net benefit for patients using the Clinic_Rad_DL model is greater than that of other models at most threshold probabilities (Figure S2).

FIGURE 4.

Receiver operating characteristic (ROC) curves of the model. The performance of the model is evaluated in the (A) training, (B) validation, and (C) test sets.

In the Cox Proportional Hazards Model, the C‐index for the training, validation, and test sets is 0.806, 0.803, and 0.819, respectively. Based on the risk scores predicted by the Cox model, patients are divided into high‐risk and low‐risk groups, with significant differences in survival times between the two groups observed in both the training and test sets (Figure 5). The multivariable Cox proportional hazards model visualizes the model using nomograms, allowing for a quick assessment of the 1, 3, and 5‐year mortality risk for patients (Figure 6).

FIGURE 5.

Kaplan–Meier (KM) survival curves of the high‐risk and low‐risk groups. Significant differences in survival probabilities between the two risk groups were observed across the (A) training, (B) validation, and (C) test sets.

FIGURE 6.

Nomogram for survival prediction. The nomogram was constructed based on a Cox proportional hazards model integrating the radiomics signature (Rad‐sign), deep learning signature (DL_sign), and clinical features. Notably, DL_sign carries the highest weight in the model.

4. Discussion

This bicentric study developed and validated a DL model and a comprehensive model combining clinical, radiomics, and DL to predict OS in GC patients who received NT, utilizing CT images obtained before and after NT. Notably, the combined model significantly outperformed the Rad model, the Clinic_Rad model, and the DL model. Our findings indicate that the combined model offers additional prognostic value beyond traditional clinicopathologic features, potentially aiding clinicians in the risk stratification of GC patients undergoing NT.

As is known to all, almost all GC patients receive routine CT examinations both at the initial diagnosis and throughout their treatment course. CT scans are frequently utilized for tumor staging and monitoring therapeutic response. Due to its widespread availability in clinical practice and its capability to distinguish between skeletal muscle and adipose tissue, CT is considered the gold standard for evaluating BC, particularly in cancer patients [21]. Previous research in this domain has predominantly employed a quantitative approach, typically utilizing a single slice at the L3 level. However, single‐slice CT measurements are prone to intrapatient variability [22]. Consequently, this study utilized CT slices of the abdominal region between the bottom part of the lung and the upper part of the pelvic bone, as this region is deemed most suitable for conducting an in‐depth analysis of BC. Furthermore, we employed TotalSegmentator, a fully automatic method for the segmentation of subcutaneous fat, visceral fat, and skeletal muscle, which offers greater accuracy and efficiency compared to the conventional manual segmentation techniques typically applied at the L3 level.

While radiomics is frequently investigated in GC patients, feature extraction typically focuses on tumors, lymph nodes, or peritumoral tissue to predict treatment response and prognosis [23]. Numerous studies have demonstrated associations between BC and various clinical outcomes in GC patients. However, the predominant focus has been on their evaluation as nutritional indices. Recent investigations, as reported by Saalfeld et al. [13] and Kroft et al. [14], have extracted radiomic features from skeletal muscle and adipose tissue, demonstrating that radiomics‐based parameters of BC analysis can predict survival outcomes in patients with hepatocellular carcinoma and pancreatic cancer. Notably, these analyses were confined to the L3 level and involved the extraction of two‐dimensional radiomic features from BC. In contrast, a recent study by Rozynek et al. has shown that automatically extracted BC radiomic features from the entire abdominal region can serve as prognostic indicators for patients with non‐small cell lung cancer. A three‐dimensional (3D) radiomic feature analysis was conducted using machine learning techniques rather than conventional statistical methods [15]. In this study, an alternative approach was adopted, utilizing delta radiomics to evaluate the relative net change in radiomic features from BC in longitudinal CT images obtained before and after NT. Unlike traditional radiomics, which rely on single time‐point CT examinations, delta‐radiomics features offer a wealth of information that can be used to identify, quantify, and potentially predict changes induced by NT throughout the therapeutic process. Our study demonstrated that the Rad model, which integrates machine learning methodologies with delta‐radiomics features, successfully predicted OS in GC patients who received NT. Nevertheless, the model's performance across the training, validation, and test sets was modest. This limitation may be attributed to the handcrafted radiomics features, which might not fully encapsulate all the valuable information present in radiological images, being constrained by the preselection of morphological and textural features.

The DL method, an advanced machine learning technique, demonstrates the capability to extract comprehensive and detailed information from raw images and to construct end‐to‐end predictive models from such data. DL is emerging as a promising approach for predicting tumor grade, treatment response, and prognosis across various cancer types [24, 25, 26]. To the best of our knowledge, the application of DL methods based on BC for predicting survival in GC patients who have received NT has not been extensively explored. Miao et al. proposed a DL model that integrates BC features with clinicopathological data, demonstrating potential for accurately predicting distant metastasis in breast cancer. Nonetheless, their analysis was limited to CT images from only two levels, T4 and T11 [27]. A recent study developed a Transformer‐based skeletal muscle deep‐learning (SMDL) model, which can accurately predict survival of GC patients and identify those at high risk of recurrence or death. However, this study employed a semi‐automatic segmentation method to analyze the skeletal muscles at the cross‐section of L3 [28]. In the current study, a 3D image encompassing subcutaneous fat, visceral fat, and skeletal muscle of the abdominal region was acquired. A hybrid DL model was developed by combining the ResNet18 network and the Transformer network, utilizing two CT scans obtained both pre‐NT and post‐NT. Furthermore, we conducted a direct comparison between the DL model and the traditional handcrafted radiomic model (Rad model), subsequently developing a fusion model. Our findings indicated that the DL model surpassed the performance of the Rad model. Additionally, the comprehensive model, which incorporated clinical features, radiomic signatures, and DL signatures, demonstrated superior performance relative to the other models. Although the AUC differences in the test set were not statistically significant, we still regard the overall performance of the combined model as superior. Evaluating model performance should involve multiple metrics (such as DCA, calibration analysis, sensitivity), as well as considerations of feasibility and practical benefits, rather than solely relying on the statistical significance of the AUC results. This proposed combined model facilitates a noninvasive and robust analysis of BC heterogeneity from medical images across multiple levels. Integrating DL features with clinical and radiological features enhanced diagnostic performance by complementing image features with additional information, thereby increasing the model's robustness [29]. Following this, a Cox regression model was established for survival analysis. Analysis based on Cox's risk scores revealed a significant disparity in survival times between high‐risk and low‐risk groups. Consequently, the findings indicated that the comprehensive model, which incorporated clinical features, radiomic signatures, and DL signatures, possesses predictive value for outcome prediction in GC patients who received NT followed by radical gastrectomy.

In current literature, skeletal muscle and adipose tissue have been recognized as significant predictors of long‐term survival in GC patients [30, 31]. The majority of prior studies have relied on single‐slice CT images to assess the areas of skeletal muscle and adipose tissue, which are of great value for evaluating nutritional status. However, this approach potentially overlooks critical information regarding BC. This study distinguishes itself from previous research with notable advantages. First, this biocentric study incorporated data from multiple hospitals, yielding optimal results. This reduces bias in our research, ensures more credibility to our findings, and confirms the efficacy of our model, all of which are beneficial for promoting the model and developing methods. Second, this study employed the automatic segmentation method to obtain a 3D image encompassing subcutaneous fat, visceral fat, and skeletal muscle across the entire abdominal region, based on two CT scans conducted pre‐NT and post‐NT. This method leverages the information obtained from abdominal CT scans of pre‐NT and post‐NT during routine diagnostic procedures, offering additional support for prognostic evaluation without increasing costs or causing harm. The innovative nature of this study is evident. Third, the proposed combined model, which integrates clinical features, radiomic signatures, and DL signatures, can noninvasively and robustly capture BC characteristics from medical images at different levels, achieving the best performance. Moreover, integrating high‐dimensional features can offer further details regarding BC, providing more sophisticated and objective information and improving the model's performance for prediction. The proposed combined model takes real‐world clinical practice into account and provides valuable guidance for clinicians, ultimately benefiting patients.

Nevertheless, this study has some limitations. Firstly, although the present study included cohorts from two centers, our models were not validated in an independent external cohort. Moreover, as the data were collected retrospectively, issues such as selection bias and hidden confounders were inevitable. Given the constraints of real‐world studies, the sample size available for analysis was restricted. The time points for undergoing CT examinations are not completely unified. The NT regimens received by these patients were not completely uniform, and it is unavoidable that biases may arise due to these differences. Therefore, large prospective datasets from diverse regions are necessary for more comprehensive validation in future studies. Furthermore, future studies may incorporate additional prognostic factors, including genomics and pathomics, to enhance the prognostic predictive capacity of the model. This integration could facilitate more precise individualized treatment decisions and improve the accuracy of prognosis predictions. In the future, cloud services, edge computing, or optimized lightweight models can be used to meet application requirements, with interfaces designed for integration into clinical workflows, along with further cost‐effectiveness analyses.

5. Conclusion

In summary, we developed and validated a DL model based on pre‐NT and post‐NT CT to predict the OS of GC patients who received NT. The comprehensive model showed better performance by combining clinical features, radiomic signatures, and DL signatures. These models have the potential to assist clinicians in the risk stratification of GC patients who received NT, thereby facilitating the formulation of risk‐adapted treatment decisions and personalized follow‐up strategies.

Author Contributions

Yingjing Zhang: conceptualization, methodology, software, investigation, formal analysis, visualization, writing – original draft, writing – review and editing. Tong Tong: data curation, resources, supervision, project administration, validation, writing – review and editing, methodology. Bowen Hou: methodology, software, conceptualization, validation, visualization, formal analysis. Tian Yu: software, writing – original draft, visualization, validation. Pengfei Su: validation, data curation, writing – review and editing. Jianchun Yu: funding acquisition, project administration, resources, supervision, data curation, conceptualization. Yakai Huang: conceptualization, data curation, supervision, project administration, resources, writing – review and editing, funding acquisition.

Funding

This study has received funding by the National High Level Hospital Clinical Research Funding (2022‐PUMCH‐B‐005).

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: Calibration curves of the predictive model. The panels show the calibration in the (A) training, (B) validation, and (C) test sets. The curves demonstrate good agreement between the predicted probabilities and the observed outcomes.

Figure S2:. Decision curve analysis (DCA) of the predictive model. The y‐axis measures the net clinical benefit, and the x‐axis represents the threshold probability. Results are shown for the (A) training, (B) validation, and (C) test sets.

Table S1:. CT Scanning Parameters.

Table S2:. Baseline feature statistics.

Table S3: Model performance metrics.

Table S4:. Delong test.

Table S5:. Hosmer‐Lemeshow test.

Data Availability Statement

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.