Prediction of response to neoadjuvant chemotherapy in advanced gastric cancer: A radiomics nomogram analysis based on CT images and clinicopathological features

Abstract

PURPOSE:

To investigate the feasibility of predicting the early response to neoadjuvant chemotherapy (NAC) in advanced gastric cancer (AGC) based on CT radiomics nomogram before treatment.

MATERIALS AND METHODS:

The clinicopathological data and pre-treatment portal venous phase CT images of 180 consecutive AGC patients who received 3 cycles of NAC are retrospectively analyzed. They are randomly divided into training set (n = 120) and validation set (n = 60) and are categorized into effective group (n = 83) and ineffective group (n = 97) according to RECIST 1.1. Clinicopathological features are compared between two groups using Chi-Squared test. CT radiomic features of region of interest (ROI) for gastric tumors are extracted, filtered and minimized to select optimal features and develop radiomics model to predict the response to NAC using Pyradiomics software. Furthermore, a nomogram model is constructed with the radiomic and clinicopathological features via logistic regression analysis. The receiver operating characteristic (ROC) curve analysis is used to evaluate model performance. Additionally, the calibration curve is used to test the agreement between prediction probability of the nomogram and actual clinical findings, and the decision curve analysis (DCA) is performed to assess the clinical usage of the nomogram model.

RESULTS:

Four optimal radiomic features are selected to construct the radiomics model with the areas under ROC curve (AUC) of 0.754 and 0.743, sensitivity of 0.732 and 0.750, specificity of 0.729 and 0.708 in the training set and validation set, respectively. The nomogram model combining the radiomic feature with 2 clinicopathological features (Lauren type and clinical stage) results in AUCs of 0.841 and 0.838, sensitivity of 0.847 and 0.804, specificity of 0.771 and 0.794 in the training set and validation set, respectively. The calibration curve generates a concordance index of 0.912 indicating good agreement of the prediction results between the nomogram model and the actual clinical observation results. DCA shows that patients can receive higher net benefits within the threshold probability range from 0 to 1.0 in the nomogram model than in the radiomics model.

CONCLUSION:

CT radiomics nomogram is a potential useful tool to assist predicting the early response to NAC for AGC patients before treatment.

Keywords

Advanced gastric cancer neoadjuvant chemotherapy radiomics nomogram Computed tomography

1 Introduction

Gastric cancer (GC) is the fifth most common cancer and the third leading cause of cancer-related deaths worldwide [1]. The five-year survival rate of patients with advanced gastric cancer (AGC) is still not optimistic because of unreasonable treatment options, and the mean survival duration is less than 1 year [2]. At present, neoadjuvant chemotherapy (NAC) has been used to improve the survival [3]. There have been several NAC regimens including ECF (epirubicin, cisplatin, and fluorouracil) that have been suggested by the National Comprehensive Cancer Network (NCCN) guidelines version 1.2017, and the effect of chemotherapy has been demonstrated in well-designed, multicenter and randomized clinical trials [4].

Currently, the clinical tumor regression was graded based on the Response Evaluation Criteria in Solid Tumors (RECIST, version 1.1), which has been considered as the gold standard for clinical evaluation of chemotherapy effect [5]. However, it is somewhat limited due to its retrospective nature after treatment. Therefore, predicting therapeutic response to NAC as early as possible has remarkable clinical benefit for AGC patients, which could optimize treatment regimen, avoid unnecessary suffering from chemotherapy toxicity, and reduce medical cost. A previous study from Sun et al has shown that low-dose CT perfusion imaging (CTPI) is a valuable tool that permits microcirculation evaluation and therefore can predict the efficacy of NAC in AGC patients [6]. However, CTPI will produce a lot of X-ray radiation, which is harmful to the human health.

With the development of computer hardware and the set-up of big database, the newly emerged study on “radiomics” is expected to be the milestone in the development of imaging field [7]. Being different from the traditional megascopic analysis based on experiences, “radiomics” is to extract plenty of quantitative features from images, and to construct signature to characterize the lesion [8, 9]. Nowadays, the radiomics has been widely used in research of lung cancer [10], liver cancer [11], breast cancer [12], and so on. However, because of the impact on the anatomical structure and physiological characteristics of the stomach, research on radiomics of the stomach is still in the exploratory stage and several studies have recently been published [13 –16]. A research has also demonstrated that the proposed pretreatment CT-based radiomics models reveal good performances with the highest AUC of 0.803 in predicting response to NAC and thus may be used to improve clinical treatment in AGC patients [17]. However, they didn’t consider clinical factors to construct the nomogram model. Therefore, our aim in this study was to establish a nomogram with visualization and interpretability in predicting the early response to NAC in AGC patients before treatment.

2 Materials and methods

2.1 Patients

This study was approved by the Institutional Review Board of our hospital, and the informed consent requirement was waved because of retrospective nature. 251 consecutive patients with gastric cancer confirmed by gastroscope biopsy from March 2017 to May 2022 were retrospectively collected. Inclusion criteria: (1) AGC (stage III IV) was confirmed according to the eighth edition of the Union for International Cancer Control/American Joint Committee on Cancer (UICC/AJCC) clinical staging system [18]; (2) Standard abdominal contrast-enhanced CT examination was performed within one week before chemotherapy; (3) Patients completed 3 cycles of NAC. Exclusion criteria: (1) Early gastric cancer (n = 6); (2) Complicated with other abdominal tumors (n = 11); (3) Incomplete clinical data (n = 18); (4) Severe artifacts on CT images (n = 16), and (5) patient unfitness for chemotherapy (n = 20). Finally, a total of 180 patients were enrolled in this study and randomly divided into training set and validation set according to the ratio of 2:1. All patients received three cycles of NAC (i.e., ECF chemotherapy regimen: epirubicin 50 mg/m2, intravenous drip, on day 1; cisplatin 60 mg/m2, intravenous drip, on day 1; 5-FU 200 mg/m2, intravenous drip, on day 1–21, being repeated every 21 days) [4], and they were divided into effective group (n = 83) and ineffective group (n = 97) according to the response evaluation criteria in solid tumors (RECIST 1.1) [5]. Clinical data of patients were recorded, including age, sex, Lauren type, clinical stage, Borrmann type, tumor location, and cancer cell differentiation.

2.2 Image acquisition and segmentation

All patients received CT examination before NAC and after three cycles of NAC. They should fast 8 hours and drink 800~1000 ml water before CT scanning. Patients were in supine position with the scanning range from the diaphragm to the inferior poles of both kidneys. First, the CT plain scanning was performed by utilizing 64 multi-slice spiral CT machine (Germany, Siemens) with slice thickness and spacing of 5 mm, tube voltage of 120 kV, tube current of 200 mA, matrix of 512×512, and pitch of 0.984:1. Second, 100 ml contract agent ioversol (320 mgI/ml) was injected via an automatic high-pressure syringe with injection speed of 3.0 ml/s. Finally, the arterial and venous-phase enhanced CT images with slice thickness and spacing of 5 mm were obtained when delaying 35 s and 65 s after the injection. The contrast-enhanced portal venous phase CT images including the largest gastric tumor slice were loaded to personal laptop from the picture archiving and communication system (PACS) (Carestream, Canada) and imported into ITK-SNAP software, and gastric tumors were manually segmented by drawing a 2D region of interest (ROI) along the tumor edge (Fig. 1) by two senior radiologists (Reader 1 and Reader 2) with 10 years and 15 years experiences in abdominal CT diagnosis, respectively, who did not know the clinicopathological results. One month later, 40 cases of gastric cancer were randomly selected from 180 cases, and the above segmentation process was repeated by Reader 1 to test the reliability and reproducibility of the segmentation. The segmented images with tumor ROI were exported and stored as DICOM format.

Fig. 1

The maps in a 58-year-old woman patient with gastric cancer. On an axial CT image on the venous phase, the manual segmentation of an irregular gastric mass is extracted via ITK-SNAP software.

2.3 Extraction and selection of features

These segmented tumor ROI images were resampled with a pixel spacing of 3.0 mm×3.0 mm×3.0 mm before features extraction to ensure the accuracy of pixel size and slice thickness. Subsequently, radiomic features, such as shape and size features, first-order features, texture and wavelet features, were extracted from the segmented images by the Pyradiomics software package based on python language (version 3.0.1, https://github.com/Radiomics/pyradiomics). First, the extracted features were preprocessed by replacing the missing values and outliers with the average values. Second, the z-score method was used to standardize large number of radiomic features according to the average and standard deviation parameters of patients in the training set [19]. Intraclass and interclass correlation coefficients (ICCs) was used to evaluate the agreement of radiomic features for intra-observer and inter-observer. An ICC greater than 0.75 represented a good agreement [20]. Final, the most useful radiomic features were selected by the least absolute shrinkage and selection operator (LASSO) regression with 10-fold cross-validation [21].

2.4 Radiomics model and nomogram model building

The linearity combination of the selected radiomic features and the corresponding weighted coefficient product constructed the radiomics signature, namely radiomics model. Clinical factors for predicting the response to NAC in AGC was identified by univariable logistic regression analysis with the following clinical candidate predictors: age, gender, Lauren type, clinical stage, Borrmann type, tumor location, cancer cell differentiation. As well, the nomogram model combining clinical factors and radiomic features was established by using multivariable logistic regression analysis.

2.5 Comparison and testing of models

The receiver operating characteristic (ROC) curve was used and the corresponding area under the curve (AUC) was reckoned for these prediction models in the training set and validation set, respectively. The AUCs were compared by using Delong test between models [22]. The calibration curve was used to verify the agreement between the prediction results of the nomogram and the actual clinical findings, and the decision curve was used to validate the value of the nomogram in clinical practice.

2.6 Statistical analysis

All statistical analyses were performed on SPSS 26.0 (SPSS Inc., Chicago, IL, USA), MedCalc version 15.11.4 (MedCalc software, Mariakerke, Belgium), and R software (version 3.5.1; http//www.R-projetc.org). A p-value < 0.05 was set, indicating a statistically significant difference. Pearson’s Chi-Squared tests were used to compare the difference in categorical variables (gender, Lauren type, clinical stage, Borrmann type, tumor location, cancer cell differentiation) between effective group and ineffective group in AGC patients. In addition, an unpaired t-test was performed to compare continuous variable (age) between the two groups. Lasso binary logistic regression was conducted using the “glmnet” package (version 2.0–13) in R software, and the nomogram, calibration curve, and decision curve were made using the “rms” package (version 5.1-2). Diagnosis efficacy was assessed using the receiver operating characteristic (ROC) curve with area under the curve (AUC), sensitivity, specificity.

3 Results

3.1 Clinical characteristics

Baseline clinical characteristics for 180 gastric cancer patients were summarized and compared in Table 1. In the univariate analysis, the response to NAC of gastric cancer had significant associations with Lauren type and clinical stage (both p < 0.05); While no significantly statistical differences were found between effective group and ineffective group patients in terms of age, gender, Borrmann type, tumor location, and cancer cell differentiation (all p > 0.05).

Table 1
Comparisons of clinicopathological characteristics between effective and ineffective groups in 180 gastric cancer patients

Clinicopathological Total Effective group Ineffective group t/ χ2 P

characteristics (n = 180) (n = 83) (n = 97) Value Value

Age, mean±SD (years) 65.66±10.94 67.28±11.75 64.28±10.06 1.845 0.067

Gender [n(%)] 0.954 0.329

Male 72(40.0) 30(36.1) 42(43.3)

Female 108(60.0) 53(63.9) 55(56.7)

Lauren type [n(%)] 20.667 0.000^*

Intestinal type 57(31.7) 40(48.2) 17(17.5)

Diffuse type 88(48.9) 28(33.7) 60(61.9)

Mixed type 35(19.4) 15(18.1) 20(20.6)

Clinical stage [n(%)] 24.347 0.000^*

III 90(50.0) 58(69.9) 32(33.0)

IV 90(50.0) 25(30.1) 65(67.0)

Borrmann type [n(%)] 1.747 0.626

I 29(16.1) 11(13.3) 18(18.6)

II 27(15.0) 12(14.5) 15(15.5)

III 76(42.2) 39(47.0) 37(38.1)

IV 48(26.7) 21(25.2) 27(27.8)

Tumor location [n(%)] 2.816 0.245

Cardia 58(32.2) 25(30.2) 33(34.1)

Body 68(37.8) 28(33.7) 40(41.2)

Antrum 54(30.0) 30(36.1) 24(24.7)

Cell differentiation [n(%)] 0.257 0.612

Low 112(62.2) 50(60.2) 62(63.9)

Moderate 68(37.8) 33(39.8) 35(36.1)

Clinicopathological	Total	Effective group	Ineffective group	t/ χ2	P
Age, mean±SD (years)	65.66±10.94	67.28±11.75	64.28±10.06	1.845	0.067
Gender [n(%)]				0.954	0.329
Male	72(40.0)	30(36.1)	42(43.3)
Female	108(60.0)	53(63.9)	55(56.7)
Lauren type [n(%)]				20.667	0.000^*
Intestinal type	57(31.7)	40(48.2)	17(17.5)
Diffuse type	88(48.9)	28(33.7)	60(61.9)
Mixed type	35(19.4)	15(18.1)	20(20.6)
Clinical stage [n(%)]				24.347	0.000^*
III	90(50.0)	58(69.9)	32(33.0)
IV	90(50.0)	25(30.1)	65(67.0)
Borrmann type [n(%)]				1.747	0.626
I	29(16.1)	11(13.3)	18(18.6)
II	27(15.0)	12(14.5)	15(15.5)
III	76(42.2)	39(47.0)	37(38.1)
IV	48(26.7)	21(25.2)	27(27.8)
Tumor location [n(%)]				2.816	0.245
Cardia	58(32.2)	25(30.2)	33(34.1)
Body	68(37.8)	28(33.7)	40(41.2)
Antrum	54(30.0)	30(36.1)	24(24.7)
Cell differentiation [n(%)]				0.257	0.612
Low	112(62.2)	50(60.2)	62(63.9)
Moderate	68(37.8)	33(39.8)	35(36.1)

NOTE. Values are mean±standard deviation, and an independent samples t-test was used to assess the difference in numeric variable (age); For the rest, values are numbers of patients with percentages in parentheses, and Pearson Chi-Squared tests were used to compare the differences in categorical variables. *p value < 0.05.

3.2 Radiomic feature selection and radscore building

A total of 464 radiomic features were initially extracted from the segmented CT images with gastric tumors ROI. The extracted features were preprocessed and standardized by z-score method and with ICC of both inter-observers and intra-observer≥0. 75 as the reliability standard, subsequently 300 features were retained. To reduce dependency and redundancy, we used LASSO in logistic regression model to reduce dimensions of these features. As a result, the following 4 features were selected.

F1: Wavelet.LL_glcm_DifferenceEntropy,

F2: Original_glszm_LargeAreaHighGrayLevelEmphasis,

F3: Wavelet.LL_glszm_ SmallAreaEmphasis,

F4: Wavelet.HH_glcm_Id.

Then, these 4 features were applied to establish the final radiomics signature model (Fig. 2), namely radscore. The radscore was defined as a score resulted by the regression coefficients of these 4 features multiplied by the value of corresponding feature (Fig. 3). The radiomics signature model was constructed by binary logistic regression method based on these 4 features to predict the response to NAC of gastric cancer using Equation (1).

Fig. 2

We used the LASSO method to select the optimized subset of radiomics feature. (A) Tuning parameter Lambda (λ) selection in the LASSO model used 10-fold cross-validation. The AUC was plotted versus log (λ). The vertical line was drawn at the optimal value where the AUC was maximum and radiomics features number was 4; (B) LASSO coefficient profiles of the retained features. A coefficient profile plot was produced against the log (λ) sequence. Vertical line was drawn at the value selected using 10-fold cross-validation, where optimal λ resulted in 4 features with nonzero coefficients. AUC, area under the curve; LASSO, least absolute shrinkage and selection operator.

Fig. 3

Four selected radiomics features and the corresponding coefficients. Radiomics signature, namely radiomics score (radscore), was calculated and constructed via summing the selected features weighted by their coefficients.

$Radscore = - 0.106 + 0.475 \times F 1 - 0.023 \times F 2 + 0.411 \times F 3 + 0.144 \times F 4$ (1)

There were significantly statistical differences in radscore between effective group and ineffective group patients in both training and validation sets (both p < 0.01) (Fig. 4).

Fig. 4

The violin plot in the training set (A) and validation set (B), showing that there are significantly statistical differences in radscore between effective group and ineffective group patients in both training and validation sets (both p < 0.001).

3.3 Development of individualized nomogram

The nomogram with visualization and interpretability was constructed by using a binary logistic regression analysis to predict the response to NAC of gastric cancer, including two clinical factors (Lauren type and clinical stage) identified by univariate analysis and the radscore (Fig. 5).

Fig. 5

The nomogram with visualization and interpretability, indicating that gastric cancer patients with Lauren intestinal type, clinical III stage, and greater radscore are tended to be predicted as the effective responder to NAC. Lauren type 1, 2 and 3 means Lauren diffuse type, mixed type and intestinal type respectively; Clinical stage 1 and 2 means IV stage and III stage respectively. NAC, neoadjuvant chemotherapy.

3.4 Performance and comparison of these predictive models

The ROC curve was used to evaluate and compare the performance of two predictive models including the radscore model and nomogram model, and the results showed that the nomogram model generated a better performance with an AUC of 0.841 (95% confidence interval [CI], 0.759–0.899) in the training set and 0.838 (95% CI, 0.720–0.920) in the validation set than the radscore model (Table 2 and Fig. 6).

Table 2
Comparisons of performance between radscore model and nomogram model in training set and validation set

Two groups AUC(95% CI) sensitivity specificity

Training set

Radscore model 0.754 (0.667–0.828) 0.732 0.729

Nomogram model 0.841 (0.759–0.899) 0.847 0.771

Validation set

Radscore model 0.743 (0.614–0.847) 0.750 0.708

Nomogram model 0.838 (0.720–0.920) 0.804 0.794

Two groups	AUC(95% CI)	sensitivity	specificity
Training set
Radscore model	0.754 (0.667–0.828)	0.732	0.729
Nomogram model	0.841 (0.759–0.899)	0.847	0.771
Validation set
Radscore model	0.743 (0.614–0.847)	0.750	0.708
Nomogram model	0.838 (0.720–0.920)	0.804	0.794

NOTE. AUC, area under the curve; CI, confidence interval.

Fig. 6

ROC curves for the radscore model in the training set (A) and validation set (B), as well as for the nomogram model in the training set (C) and validation set (D), indicating that classifiers’ performance of the nomogram model in predicting the response to NAC of gastric cancer is better than that of the radscore model. ROC, receiver operating characteristic; NAC, neoadjuvant chemotherapy.

Delong-test results indicated that there was a significantly statistical difference between predictive performance of the nomogram model and that of the radscore model (AUC = 0.841 vs. 0.754, p = 0.001) in the training set. As well, there was a significantly statistical difference in the validation set (AUC = 0.838 vs. 0.743, p = 0.002).

The calibration curve of the nomogram model showed a good agreement with an index of concordance of 0.912 (95% confidence interval [CI], 0.841–0.983) between prediction probability and actual clinical findings (Fig. 7a). The decision curve in the validation set proved that the nomogram model was an useful tool, and indicated that if the threshold probability of a patient with AGC was within the whole range of 0.0–1.0, using the nomogram to predict the response to NAC added more net benefit to make the decision of whether to undergo treatment than the radscore (Fig. 7b).

Fig. 7

Calibration curve (A) and decision curve analysis (B) for the nomogram in the validation set. In the calibration plot, X-axis is nomogram-predicted probability of effective response to NAC. Y-axis is true probability. Dotted diagonal line represents an ideal standard curve; solid line represents the prediction calibration curve of the nomogram. A good agreement is observed between nomogram prediction probability and actual clinical findings in this calibration plot (index of concordance is 0.912). The decision curve plot indicated that if the threshold probability of a patient with AGC was within the whole range of 0.0–1.0, using the nomogram to predict the response to NAC added more net benefit to make the decision of whether to undergo treatment than the radscore. NAC, neoadjuvant chemotherapy; AGC, advanced gastric cancer.

4 Discussion

In this study, the response to NAC of gastric cancer had significant associations with Lauren type and clinical stage in clinical characteristics. In radiomics analysis, a total of 4 radiomic features were selected to establish the radiomics signature model. Finally, we constructed a CT radiomics-based nomogram that incorporated both the radiomics signature and clinical factors to predict the response to NAC of gastric cancer prior to chemotherapy, which yielded a better performance with an AUC of 0.841 in the training set and 0.838 in the validation set than radiomics signature model. As well, this study demonstrated that the nomogram with visualization and interpretability was an easy-to-use tool for individualized decision making by quantifying the prediction model.

In clinic, NAC has become a popular treatment modality for AGC worldwide; However, the therapeutic response of AGC to NAC is highly heterogeneous, and the prognosis of patients who have good responses is significantly better than that of patients with poor responses [23]. At present, there is no reliable and effective method to predict the curative effect of NAC for AGC, which leads to the failure of NAC in some patients due to the insensitivity, and some patients even lose the chance of radical surgery when their disease progresses during chemotherapy. Therefore, it is of great significance to develop a model to accurately predict the curative effect of NAC before treatment for patients with AGC. Of course, the response evaluation criteria in solid tumors (RECIST 1.1) has been considered the standard for clinical therapeutic evaluation of NAC in gastric cancer, but its application is somewhat hysteretic due to its retrospective nature by calculating the tumor regression rate in size between before and after treatment [5]. In contrast, predicting therapeutic response to NAC as early as possible has remarkable clinical benefit for AGC patients, which could optimize treatment regimen, avoid unnecessary suffering from NAC toxicity, and reduce medical cost.

In a recent study, it has been found that pretreatment dynamic contrast-enhanced MRI (DCE-MRI) quantitative parameters Ktrans, kep, ve, and intravoxel incoherent motion diffusion-weighted imaging (IVIM-DWI) quantitative parameter D value are independent predictors of NAC response for AGC patients [24]. Although their combined predictive model, which consisted of DCE-MRI and IVIM-DWI, showed the best prediction performance with an area under the curve (AUC) of 0.922. However, due to respiratory movement and gastrointestinal motility, these functional MRI techniques with longer scanning time than CT scanning were rarely used in gastric research. So the inferior MRI image quality in gastric cancer which might have an influence on the result or conclusion of their study needed to be further improved. On the contrary, the CT imaging of gastric cancer in this study has better image quality and resolution.

In addition, with the rapid development of artificial intelligence techniques and the set-up of big database, researchers have tried to break through the traditional megascopic image analysis based on experiences and seek a noninvasive and objective method to quantitatively characterize the lesion and disclose the heterogeneity of tumor. The newly emerged radiomics which could extract plenty of quantitative features from CT or MRI images and construct radiomics signature and nomogram has been widely used to predict the response to NAC or chemo-radiotherapy (CRT) of breast tumor, rectal cancer, lung cancer and so on [25 –27]. In a preliminary study, Yoshida et al. [25] found that the best diagnostic performance was achieved when both first and second order texture features based on pretreatment DCE-MRI radiomics with clinical information and subjective radiological findings were used (AUC = 0.77), and they drew a conclusion that pretreatment DCE-MRI radiomics could improve the prediction of pathological complete response (pCR) to NAC in breast cancer patients. In another study, Bonomo et al. [26] demonstrated CT-based radiomics may be a feasible and reproducible biomarker for prediction of response after neoadjuvant CRT in patients with locally advanced rectal cancer, yielding a prediction performance with AUC of 0.63. Khorrami et al. [27] disclosed radiomic texture features extracted within and around the lung tumor on CT images may be used to predict pathologic response to neoadjuvant CRT in resectable stage III non-small cell lung cancer patients, yielding an AUC of 0.90 within the training set and a corresponding AUC of 0.86 within the validation set.

However, research on CT radiomics of predicting the response to NAC in AGC patients is still in the exploratory stage and few studies have been published [28 –30]. Song et al. [28] investigated the ability of the CT-based radiomics models for pretreatment prediction of the response to NAC in AGC patients, and showed the prediction performance with the highest AUC of 0.790 in the training cohort and the corresponding AUCs of 0.784 and 0.803 in the internal and external validation cohorts, respectively. However, they didn’t build a nomogram model based on radiomic features in combination with clinicopathological factors, so the predictive ability of their radiomics models was not as good as that of our nomogram model. In a similar study, Xie et al. [29] also proposed and validated a practical prediction method of pathological regression following NAC based on single contrast-enhanced computed tomography (CECT) radiomics models in AGC patients across different hospitals. In addition, Chen et al. [30] developed and validated a nomogram for predicting survival in AGC patients after NAC and radical surgery by integrating multiple clinical factors and using a Cox regression model to determine their impact on survival, and the C-index and AUC of the established nomogram prediction model was 0.785 and 0.736, respectively. In comparison with our research, they didn’t analyze and extract radiomic features of the gastric tumors ROI on the CT images, which resulted in the unsatisfactory prediction performance of their nomogram model.

Through the interpretation of the nomogram in this study, it could be found that the better response to NAC in AGC patients related to Lauren intestinal type and clinical III stage. These results were consistent with previous studies [31, 32]. In addition, we extracted radiomic features based on single 2D ROI of gastric tumor in CT image slices. A research has suggested that 2D CT annotations might be a preferred choice in gastric cancer radiomics studies than 3D because 3D annotations might bring more noise [33]. The LASSO method confirmed that 4 radiomic texture features were related to the response to NAC of AGC and were used as independent predictors. Then, instead of single factor analysis, we built and verified a nomogram for AGC patients based on the combined radiomics signature with clinical factors to predict the response to NAC. Up to now, our nomogram is the most promising noninvasive and quantitative prediction tool to identify the response to NAC of AGC prior to treatment.

There are also some limitations in this study. First, this is a retrospective study in our single-center hospital and external validation of multi-center cases can be required to evaluate the suitability of this nomogram model. Second, only venous phase CT images were used to extract radiomic features as it is difficult to identify tumor margins exactly on unenhanced or arterial phase CT images, although they may contain some useful information. Third, in this study, due to the short follow-up time, we didn’t perform the correlation analysis study between survival of patients and radiomic features of gastric tumors. Perhaps, this had more important clinical significance.

In conclusion, this nomogram merging radiomic features of gastric tumor, Lauren type, and clinical stage showed good performance in predicting the response to NAC of AGC before chemotherapy, which was a helpful and easy-to-use tool for individualized decision making in clinical treatment strategy management because of its visualization and interpretability.

Footnotes

Acknowledgments

This study was funded by Top Talent Support Program for young and middle-aged people of Wuxi Health Committee in China (grant number: HB2020046), and clinical research and translational medicine research program of affiliated hospital of Jiangnan university in China (grant number: LCYJ202209).

References

Siegel

R.L.

, Miller

K.D.

, Jemal

Cancer statistics, CA Cancer J Clin 69 (2019), 7–34.

Thrift

A.P.

, El-Serag

H.B.

Burden of gastric cancer, Clin Gastroenterol Hepatol 18 (2020), 534–542.

Jiang

, Zeng

, Zhang

et al., Lymph node ratio is a prospective prognostic indicator for locally advanced gastric cancer patients after neoadjuvant chemotherapy, World J Surg Oncol 20 (2022), 261.

Al-Batranh

S.E.

, Homann

, Pauligk

et al., Perioperative chemotherapy with fluorouracil plus leucovorin, oxaliplatin, and docetaxel versus fluorouracil or capecitabine plus cisplatin and epirubicin for locally advanced, resectable gastric or gastro-oesophageal junction adenocarcinoma (FLOT4): A randomised, phase 2/3 trial, Lancet 393 (2019), 1948–1957.

Eisenhauer

, Therasse

, Bogaerls

et al., New response evaluation criteria in solid tumors: revised RECIST guideline (version 1.1), Eur J Cancer 45 (2009), 228–247.

Sun

Z.Q.

, Cheng

X.F.

, Ge

Y.X.

et al., An application study of low-dose computed tomography perfusion imaging for evaluation of the efficacy of neoadjuvant chemotherapy for advanced gastric adenocarcinoma, Gastric Cancer 21 (2018), 413–420.

Guiot

, Vaidyanathan

, Deprez

et al., A review in radiomics: Making personalized medicine a reality via routine imaging, Med Res Rev 42 (2022), 426–440.

Sah

B.R.

, Owczarczyk

, Siddique

et al., Radiomics in esophageal and gastric cancer, Abdom Radiol 44 (2019), 2048–2058.

Lambin

, Leijenaar

R.T.H.

, Deist

T.M.

et al., Radiomics: the bridge between medical imaging and personalized medicine, Nat Rev Clin Oncol 14 (2017), 749–762.

10.

Hou

K.Y.

, Chen

J.R.

, Wang

Y.C.

et al., Radiomics-based deep learning prediction of overall survival in non-small-cell lung cancer using contrast-enhanced computed tomography, Cancers (Basel) 14 (2022), 3798.

11.

, Cao

, Du

et al., Radiomics for the detection of microvascular invasion in hepatocellular carcinoma, World J Gastroenterol 28 (2022), 2176–2183.

12.

Pesapane

, Agazzi

G.M.

, Rotili

et al., Prediction of the pathological response to neoadjuvant chemotherapy in breast cancer patients with MRI-radiomics: A systematic review and Meta-analysis, Curr Probl Cancer 46 (2022), 100883.

13.

Sun

, Jin

, Zhang

et al., Preoperative prediction for Lauren type of gastric cancer: A radiomics nomogram analysis based on CT images and clinical features, J Xray Sci Technol 29 (2021), 675–686.

14.

Liang

, Wu

, Liu

et al., A multicenter study on the preoperative prediction of gastric cancer microsatellite instability status based on computed tomography radiomics, Abdom Radiol 47 (2022), 2036–2045.

15.

Chen

, Zhang

, Liu

et al., Radiomics in precision medicine for gastric cancer: opportunities and challenges, Eur Radiol 32 (2022), 5852–5868.

16.

Wang

, Lv

, Xue

et al., Novel CT based clinical nomogram comparable to radiomics model for identification of occult peritoneal metastasis in advanced gastric cancer, Eur J Surg Oncol (2022), DOI: 10.1016/j.ejso.2022.06.034.

17.

Song

, Cui

, Ren

et al., CT-based radiomics analysis in the prediction of response to neoadjuvant chemotherapy in locally advanced gastric cancer: A dual-center study, Radiother Oncol 171 (2022), 155–163.

18.

Amin

M.B.

, Greene

F.L.

, Edge

S.B.

et al., The eighth edition AJCC cancer staging manual: Continuing to build a bridge from a population-based to a more “personalized” approach to cancer staging, CA Cancer J Clin 67 (2017), 93–99.

19.

Beaumont

, Iannessi

, Bertrand

A.S.

et al., Harmonization of radiomic feature distributions: impact on classification of hepatic tissue in CT imaging, Eur Radiol 31 (2021), 6059–6068.

20.

Ionan

A.C.

, Polley

M.C.

, McShane

L.M.

et al., Comon of confidence interval methods for an intra-class correlation coefficient (ICC), Bmc Med Res Methodol 14 (2014), 121–paris.

21.

Zhao

, Li

, Zhang

et al., Radiomics analysis of CT imaging improves preoperative prediction of cervical lymph node metastasis in laryngeal squamous cell carcinoma, Eur Radiol (2022). DOI: 10.1007/s00330-022-09051-4.

22.

DeLong

E.R.

, DeLong

D.M.

, Clarke-Pearson

D.L.

Comparing the areas under two or more correlated receiver operating characteristic curves: A nonparametric approach, Biometrics 44 (1988), 837–845.

23.

Huang

, Li

, Liu

et al., Enhanced CT-based radiomics predicts pathological complete response after neoadjuvant chemotherapy for advanced adenocarcinoma of the esophagogastric junction: a two-center study, Insights Imaging 13 (2022), 134.

24.

Zhu

, Jiang

, Wang

et al., Quantitative dynamic-enhanced MRI and intravoxel incoherent motion diffusion-weighted imaging for prediction of the pathological response to neoadjuvant chemotherapy and the prognosis in locally advanced gastric cancer, Front Oncol 12 (2022), 841460.

25.

Yoshida

, Kawashima

, Kannon

et al., Prediction of pathological complete response to neoadjuvant chemotherapy in breast cancer using radiomics of pretreatment dynamic contrast-enhanced MRI, Magn Reson Imaging 92 (2022), 19–25.

26.

Bonomo

, Socarras Fernandez

, Thorwarth

, et al., Simulation CT-based radiomics for prediction of response after neoadjuvant chemo-radiotherapy in patients with locally advanced rectal cancer, Radiat Oncol 17 (2022), 84.

27.

Khorrami

, Jain

, Bera

et al., Predicting pathologic response to neoadjuvant chemoradiation in resectable stage III non-small cell lung cancer patients using computed tomography radiomic features, Lung Cancer 135 (2019), 1–9.

28.

Song

, Cui

, Ren

et al., CT-based radiomics analysis in the prediction of response to neoadjuvant chemotherapy in locally advanced gastric cancer: A dual-center study, Radiother Oncol 171 (2022), 155–163.

29.

Xie

, Cui

, Zhang

et al., Pretreatment contrast-enhanced computed tomography radiomics for prediction of pathological regression following neoadjuvant chemotherapy in locally advanced gastric cancer: A preliminary multicenter study, Front Oncol 11 (2021), 770758.

30.

Chen

, Liu

, Xiao

et al., Nomogram for predicting survival in advanced gastric cancer after neoadjuvant chemotherapy and radical surgery, Gastroenterol Res Pract 2021 (2021), 2923700.

31.

Chen

, Xu

, Li

Y.L.

et al., CT-based radiomics showing generalization to predict tumor regression grade for advanced gastric cancer treated with neoadjuvant chemotherapy, Front Oncol 12 (2022), 758863.

32.

Chen

, Yuan

, Wang

et al., Evaluation of dual-energy CT derived radiomics signatures in predicting outcomes in patients with advanced gastric cancer after neoadjuvant chemotherapy, Eur J Surg Oncol 48 (2022), 339–347.

33.

Meng

, Dong

, Chen

et al., 2D and 3D CT Radiomic features performance comparison in characterization of gastric cancer: a multi-center study, IEEE J Biomed Health Inform 25 (2021), 755–763.