A CT-based deep learning radiomics nomogram for predicting the response to neoadjuvant chemotherapy in patients with locally advanced gastric cancer: A multicenter cohort study.

Background: Accurate prediction of treatment response to neoadjuvant chemotherapy (NACT) in individual patients with locally advanced gastric cancer (LAGC) is essential for personalized medicine. We aimed to develop and validate a deep learning radiomics nomogram (DLRN) based on pretreatment contrast-enhanced computed tomography (CT) images and clinical features to predict the response to NACT in patients with LAGC.
Methods: 719 patients with LAGC were retrospectively recruited from four Chinese hospitals between Dec 1st, 2014 and Nov 30th, 2020. The training cohort and internal validation cohort (IVC), comprising 243 and 103 patients, respectively, were randomly selected from center I; the external validation cohort1 (EVC1) comprised 207 patients from center II; and EVC2 comprised 166 patients from another two hospitals. Two imaging signatures, reflecting the phenotypes of the deep learning and handcrafted radiomics features, were constructed from the pretreatment portal venous-phase CT images. A four-step procedure, including reproducibility evaluation, the univariable analysis, the LASSO method, and the multivariable logistic regression analysis, was applied for feature selection and signature building. The integrated DLRN was then developed for the added value of the imaging signatures to independent clinicopathological factors for predicting the response to NACT. The prediction performance was assessed with respect to discrimination, calibration, and clinical usefulness. Kaplan-Meier survival curves based on the DLRN were used to estimate the disease-free survival (DFS) in the follow-up cohort (n = 300). Findings: The DLRN showed satisfactory discrimination of good response to NACT and yielded the areas under the receiver operating curve (AUCs) of 0.829 (95% CI, 0.739-0.920), 0.804 (95% CI, 0.732-0.877), and 0.827 (95% CI, 0.755-0.900) in the internal and two external validation cohorts, respectively, with good calibration in all cohorts (p > 0.05). Furthermore, the DLRN performed significantly better than the clinical model (p < 0.001). Decision curve analysis confirmed that the DLRN was clinically useful. Besides, DLRN was significantly associated with the DFS of patients with LAGC (p < 0.05). Interpretation: A deep learning-based radiomics nomogram exhibited a promising performance for predicting therapeutic response and clinical outcomes in patients with LAGC, which could provide valuable information for individualized treatment.

Evidence before this study

We searched PubMed and Web of Science with the terms “(radiomics OR deep learning) AND (predict OR prediction) AND (response OR non-response) AND neoadjuvant chemotherapy AND gastric cancer” for papers published from database inception to Dec 31, 2021, with no language restrictions. We found seven original studies that applied radiomics analysis to predict response to neoadjuvant chemotherapy, displaying moderate performance with areas under the receiver operating curve of 0.70-0.75.

Added value of this study

To our knowledge, no report has investigated the potential benefits of combining deep learning and radiomics to enhance prediction performance. Our deep learning radiomics model, combining clinical factors with deep learning and radiomics features, outperformed the clinical model by discrimination and calibration in the large multicenter cohorts. Additionally, higher deep learning radiomics nomogram scores were significantly associated with improved disease-free survival, providing useful complementary information for individualized treatment.

Implications of all the available evidence

Our findings show that the deep learning radiomics model has a better predictive performance of therapeutic response and prognosis before treatment than the clinical model in patients with locally advanced gastric cancer, which could provide a novel tool for guiding neoadjuvant therapy and individualized treatment strategies. Future prospective multicenter studies with larger cohorts, ideally randomized controlled trials, are needed to validate our findings.

Alt-text: Unlabelled box

Introduction

Gastric cancer (GC) remains the fourth leading cause of cancer-related deaths worldwide and one of the most common malignancies in Asia. The majority of patients are diagnosed at an advanced stage with a poor prognosis, the 5-year overall survival rate is 30–40% after curative resection. Relapse-related death remains a challenge for curative treatment. Currently, several strategies have evolved to improve survival. Specifically, neoadjuvant chemotherapy (NACT) has demonstrated favorable results with improved R0 resection rate and prognosis in patients with local advanced GC (LAGC)., Nevertheless, not all patients benefit from neoadjuvant therapies, due to the nature of tumor heterogeneity. Clinically, histopathological examination is still the gold standard method for response evaluation; however, it is only accessible after surgery, delaying the timely adjustment of therapy. Therefore, a reliable approach to early and individual predictive response is an urgent requisite for personalized treatment for patients with LAGC.

Several studies have proposed methods to identify good responders to NACT and found that some clinicopathological and molecular biomarkers, including tumor location, tumor differentiation, tumor-infiltrating lymphocytes (TILs), and death-associated protein-3 (DAP-3), are valuable for curative prediction.5, 6, 7, 8 However, none provide quantifiable risk measures, and the accuracy is limited. By contrast, some quantitative features reflective of tumor pathophysiological or metabolic characteristics, derived from different imaging modalities, such as low-dose computed tomography (CT) perfusion imaging and diffusion kurtosis imaging (DKI), had been shown to be associated with the response towards NACT.9, 10, 11, 12 However, these methods exhibit technical differences or may focus on the mean values, discarding a wealth of information on spatial tumor heterogeneity.

Recently, radiomics—converting medical images into mineable high-throughput quantitative features—has garnered increasing attention., Emerging evidence has confirmed that radiomics is an innovative strategy for tumor diagnosis and prediction of GC.15, 16, 17 Indeed, several radiomics models have been established to predict response to NACT in patients with LAGC.18, 19, 20 However, the clinical utility of these models is limited by the potential risks of overfitting due to small sample sizes and lack of independent validation. Moreover, the introduction of deep learning (DL) features enables radiomics to obtain intricate structures related to specific tasks, which showed excellent performance in tumor characterization and prognostic prediction in terms of esophageal, gastric, rectal and nasopharyngeal cancers.21, 22, 23, 24 To the best of our knowledge, to date, no study has included the association of deep learning radiomics to the therapeutic response prediction of NACT in patients with LAGC.

Therefore, we aimed to develop and validate a deep learning radiomics nomogram (DLRN) for early prediction of good response prior to the administration of NACT in a large-scale multicenter patient cohort. We also explored the ability of DLRN to predict prognosis in a subset of patients as the secondary goal of this study.

Methods

Patients

Ethics approvals were obtained from the Institutional Review Board of all participating centers, and the requirement for informed consent was waived due to the retrospective nature of this multicenter study.

A total of 719 patients with histologically confirmed GC at a locally advanced stage (cT2-4N0/+M0) who received NACT followed by gastrectomy from four independent hospitals in China were recruited. The detailed inclusion/exclusion criteria and enrolment process are illustrated in Appendix E1 and Fig. S1.

Subsequently, 346 patients with LAGC between Jan 1st, 2017 and Nov 30th, 2020 from center I were reviewed and randomly divided into a training cohort (TC, n = 243) and an internal validation cohort (IVC, n = 103) at a ratio of 7:3. Furthermore, 207 patients from center II (Yunnan Cancer Hospital) between Dec 31st, 2014 and Aug 31st, 2020 were collected as external validation cohort 1 (EVC1). We also curated an external validation cohort 2 (EVC2) comprising 166 patients from center III (The Sixth Affiliated Hospital of Sun Yat-sen University, n = 130) and center IV (Sichuan Provincial Cancer Hospital, n = 36) between Jan 1st,2015 and April 30th, 2020. The sample size estimation is shown in Appendix E2.

All baseline clinical characteristics, including age, sex, body mass index (BMI), tumor differentiation, CEA, CA199, and clinical T(cT) and N(cN) stages, according to the 8th AJCC TNM staging system, were retrieved from medical records (Table S1). In addition, a follow-up cohort (n = 300) was used for survival analysis in LAGC patients (Appendix E3).

Neoadjuvant chemotherapy protocols and response assessment

All enrolled patients received 2–4 cycles of NACT before surgery (i.e., SOX regimen: Oxaliplatin 130 mg/m2, intravenous drip, on day 1; and S-1 80 mg/m2, orally, on days 1–14; repeated every 21 days), and gastrectomy was performed within 2 weeks after NACT. The response to NACT was assessed by consensus of two subspecialty pathologists with 20 years and 10 years of experience in gastrointestinal cancer, respectively, who were blinded to the study data. The tumor regression grade (TRG) criteria were selected according to the recent National Comprehensive Cancer Network (NCCN) guideline (version 4, 2021). Briefly, the lack of any viable cancer cell in the primary lesion or lymph nodes was defined as complete response (TRG 0), while TRG 3 was defined as no evident tumor regression with extensive residual cancer. The presence of single cells or rare small groups of cancer cells was classified TRG 1, and evident tumor regression but more than single cells or rare small groups cells was defined as TRG 2. Furthermore, patients with TRGs 0-1 were considered good response (GR), while those with TRGs 2–3 were defined as poor response (PR) (Fig. S2).

CT examination and image preprocessing

Figure 1 shows the workflow of this study. All patients underwent enhanced CT examination within two weeks before NACT. Portal venous-phase CT images were retrieved from Picture Archiving and Communication Systems (PACS) for further evaluation. Details regarding the CT image acquisition settings and image processing are shown in Appendix E4 and Table S2. Tumor regions of interest (ROIs) were manually delineated by reader 1 (K.W., with 8 years of experience in abdominal CT interpretation) on the central slice of CT images with the largest tumor, according to relevant literature, using the MITK software (version 2013.12.0; http://www.mitk.org/). After one month, 50 patients were randomly chosen and segmented again by readers 1 and 2 (Z.L., with 15 years of experience in abdominal CT interpretation) to evaluate the inter-/intra-observer reproducibility of radiomics features.

Figure 1

Workflow of the study. Workflow of deep learning radiomics nomogram (DLRN) modeling for good response (GR) prediction in patients with locally advanced gastric cancer (LAGC). CT, computed tomography.

DL radiomics feature extraction

A total of 1125 handcrafted features and 1024 DL features were extracted from each ROI based on portal venous-phase CT images to quantify the tumor phenotype (Appendix E5). Herein, we adapted the DenseNet-121 architecture to extract DL features. The handcrafted features included shape, first order statistics, texture, and transformation features.

Deep learning and handcrafted signature building

As shown in Appendix E6, feature selection and signature building were performed in TC according to the following steps: (i) intra-/interclass correlation coefficients (ICCs) were calculated to explore the feature reproducibility; (ii) reservation of features with p < 0.01 using the univariable analysis, independent t-test or Mann–Whitney U test, as appropriate; (iii) the least absolute shrinkage and selection operator (LASSO) method was used to select the significant features; (iv) signature building with multivariable logistic regression analysis. Consequently, two types of signatures, handcrafted and DL, reflecting the different phenotypic characteristics of the tumors, were built as the predictors of GR status. The signatures combining the DL and handcrafted features were also constructed.

DLRN construction

In the TC, univariable analysis was carried out to select statistically significant clinicopathological variables (p < 0.05). Multivariable logistic regression was conducted to build the DLRN by combining the handcraft and DL signatures, as well as the significant clinicopathological factors. Additionally, a clinical model containing only clinicopathological variables was built for comparison.

Performance evaluation

The performances of all established models were measured using the receiver operator characteristic (ROC) analysis, and the area under the ROC curve (AUC) was calculated and compared among the cohorts using the DeLong test. Also, the predictive accuracy, sensitivity, and specificity were measured. The calibration curve was plotted to assess the calibration of all the models in the training and validation cohorts via bootstrapping with 1000 resamples, accompanied by the Hosmer-Lemeshow goodness-of-fit test.

Net reclassification index (NRI) and integrated discrimination improvement (IDI) were calculated to compare the performance between DLRN and the clinical model. The confusion matrix of DLRN was also depicted. Additionally, stratified analyses were performed on all patients’ clinicopathological characteristics and CT protocol. The clinical usefulness of the models was evaluated with decision curve analysis (DCA) by quantifying the net benefit at various threshold probabilities. Besides, the association between DLRN score and disease-free survival (DFS) was evaluated in the follow-up cohort using Kaplan-Meier curve.

Statistical analysis

The differences in the clinicopathological characteristics between patients in different groups or cohorts were compared using the independent t-test or Mann–Whitney U test for continuous variables, and Fisher's exact test or chi-Squared test for categorical variables, as appropriate. The DFS probabilities were evaluated by Kaplan-Meier survival analysis and the log-rank test. The optimal cutoff value was determined using the maximally selected rank statistics method, and patients were classified into high-risk or low-risk groups. The univariable and multivariable analyses with Cox proportional hazards regression, with backward stepwise elimination and Akaike information criteria (AIC), were used to construct the models for DFS prediction. The proportional hazards assumption of models was verified by examining the scaled Schoenfeld residual test. All statistical analyses were performed using R software version (version 3.6.3, http://www.R-project.org, Appendix E7). A two-sided p < 0.05 was considered statistically significant.

Role of the funding source

The funding source had no involvement in study design, data collection, data analysis, or manuscript preparation or approval. All authors had full access to all the data and approved the final manuscript for submission.

Results

Baseline information

The baseline characteristics of all 719 patients with LAGC are summarized in Tables 1 and S1. Patients in the four cohorts were balanced for the efficacy of NACT, with the GR rate of 23.9% and 23.3% for TC and IVC, and 17.9% and 24.7% for EVC1 and EVC2, respectively. No significant difference was detected in the age, BMI, sex, differentiation, pre-NACT CEA, CA199 and tumor location, as well as cN stage between the GR and PR groups in all four cohorts (p > 0.05). Moreover, cT stage showed significant differences between the two groups in TC and EVC2 (p < 0.05), and was then constructed for the clinical model.

Table 1

Clinicopathological characteristics of patients with LAGC in the training and validation cohorts.

Characteristics	Training cohort			Internal Validation cohort			External Validation cohort 1			External Validation cohort 2
	GR (n = 58)	PR (n = 185)	P value	GR (n = 24)	PR (n = 79)	P value	GR (n = 37)	PR (n = 170)	P value	GR (n = 41)	PR (n = 125)	P value
Age(y), median (IQR)	61.0(54.3–66.0)	60.0(53.0–65.0)	0.581	61.5(52.5–65.3)	60.0(52.5–64.0)	0.821	58.0(51.0–66.0)	55.0(48.0–63.0)	0.115	61.0(53.0–67.0)	60.0(51.0–64.0)	0.409
BMI, median (IQR)	22.9(20.6–25.0)	22.8(20.4–24.8)	0.989	21.9(20.1–24.7)	22.3(20.4–25.2)	0.676	21.6(20.3–22.5)	21.6(20.0–24.0)	0.657	22.4(19.9–23.2)	22.5(20.2–24.1)	0.401
Sex, No. (%)			1.000			0.728			0.070			0.955
Female	13(22.4%)	40(21.6%)		2(8.3%)	11(13.9%)		6(16.2%)	56(32.9%)		9(22.0%)	30(24.0%)
Male	45(77.6%)	145(78.4%)		22(91.7%)	68(86.1%)		31(83.8%)	114(67.1%)		32(78.0%)	95(76.0%)
Differentiation (%)			0.164			0.106			0.440			1.000
well	1(1.7%)	6(3.2%)		3(12.5%)	2(2.5%)		0(0.0%)	4(2.3%)		2(4.9%)	5(4.0%)
Moderately	9(15.5%)	49(26.5%)		8(33.3%)	23(29.1%)		9(24.3%)	28(16.5%)		11(26.8%)	33(26.4%)
Poorly	48(82.8%)	130(70.3%)		13(54.2%)	54(68.4%)		28(75.7%)	138(81.2%)		28(68.3%)	87(69.6%)
Pre-NACT CEA (%)			0.060			0.290			1.000			1.000
≤5(Normal)	41(70.7%)	103(55.7%)		11(45.8%)	48(60.8%)		18(48.6%)	81(47.6%)		21(51.2%)	66(52.8%)
>5(Abnormal)	17(29.3%)	82(44.3%)		13(54.2%)	31(39.2%)		19(51.4%)	89(52.4%)		20(48.8%)	59(47.2%)
Pre-NACT CA199(%)			0.325			0.074			0.716			0.873
≤20(Normal)	47(81.0%)	136(73.5%)		16(66.7%)	67(84.8%)		28(75.7%)	136(80.0%)		32(78.0%)	94(75.2%)
>20(Abnormal)	11(19.0%)	49(26.5%)		8(33.3%)	12(15.2%)		9(24.3%)	34(20.0%)		9(22.0%)	31(25.0%)
Locations, No. (%)			0.185			0.162			0.457			0.093
Cardia	26(44.8%)	84(45.4%)		10(41.7%)	44(55.7%)		7(18.9%)	17(10.0%)		14(34.1%)	41(32.8%)
Gastric body	11(19.0%)	55(29.7%)		5(20.8%)	18(22.8%)		11(29.7%)	54(31.8%)		6(14.6%)	40(32.0%)
Gastric antrum	18(31.0%)	35(18.9%)		9(37.5%)	13(16.5%)		19(51.4%)	97(57.1%)		20(48.8%)	40(32.0%)
Whole stomach	3(5.2%)	11(5.9%)		0(0.00%)	4(5.1%)		0(0.00%)	2(1.2%)		1(2.4%)	4(3.20%)
Clinical T stage (%)			0.001*			0.200			0.472			0.005*
T2	0(0.0%)	1(0.5%)		2(8.3%)	1(1.3%)		0(0.0%)	8(4.71%)		4(9.8%)	5(4.0%)
T3	33(56.7%)	60(32.4%)		8(33.3%)	29(36.7%)		12(32.4%)	40(23.5%)		29(70.7%)	64(51.2%)
T4a	25(43.1%)	109(58.9%)		13(54.2%)	48(60.8%)		20(54.1%)	101(59.4%)		4(9.8%)	44(35.2%)
T4b	0(0.0%)	15(8.1%)		1(1.3%)	1(4.2%)		5(13.5%)	21(12.4%)		4(9.8%)	12(9.6%)
Clinical N stage (%)			0.057			0.509			0.081			0.937
N0	11(19.0%)	17(9.2%)		4(16.7%)	11(13.9%)		4(10.8%)	23(13.5%)		3(7.3%)	7(5.6%)
N1	21(36.2%)	51(27.6%)		10(41.7%)	22(27.8%)		14(37.8%)	30(17.6%)		15(36.6%)	42(33.6%)
N2	16(27.6%)	66(35.7%)		6(25.0%)	24(30.4%)		9(24.3%)	52(30.6%)		16(39.0%)	53(42.4%)
N3	10(17.2%)	51(27.6%)		4(16.7%)	22(27.8%)		10(27.0%)	65(38.2%)		7(17.1%)	23(18.4%)

Abbreviations: BMI, body mass index; NACT, neoadjuvant chemotherapy; CEA, carcinoembryonic antigen;.

NOTE: Chi-squared or Fisher's exact tests, were used to compare the differences in categorical variables, whereas student t or Mann-Whitney U test was used to compare the differences in continuous variables, as appropriate. *P < 0.05.

Radiomics and DL signatures validation

Finally, 10 and 18 features were selected to build the radiomics signature and DL signature, respectively. The detailed process and the selected features for the construction of the Rad-score are described in Appendix E8 and Table S3. As shown in Tables 2 and S4, the DL-based signature achieved better predictive performance than the corresponding handcrafted signatures, with AUCs of 0.808 (95% CI, 0.746–0.870) and 0.806 (95% CI, 0.705–0.907) in the TC and IVC, respectively. Then, the moderate performances of the DL-based signature were validated, with AUCs of 0.720 (95% CI, 0.631–0.808) and 0.734 (95% CI, 0.642–0.827) in the EVC1 and EVC2, respectively, although slightly lower than the corresponding handcrafted signatures. Furthermore, the signature combining both handcrafted features and DL features achieved an improved performance than either of them alone.

Performance and validation of DLRN

In the TC, the handcraft-based signature, DL-based signature, and cT stages were independent factors for GR prediction using backward stepwise multivariable analysis with the lowest AIC criteria (Table S5 and Fig. S3), and were then combined into DLRN (Figure 2A). These selected imaging features are weakly correlated or uncorrelated (Fig. S4).

Figure 2

Deep learning radiomics nomogram (DLRN) and its performance. (A) DLRN with the handcrafted and deep learning signatures and clinical T stage. (B) Box plots showing patterns of correlation between therapeutic response and DLRN score for in the TC, IVC, EVC1, and EVC2, respectively. (C) Calibration curves of DLRN in all the four cohorts. (D) Decision curve analysis for DLRN, deep learning signature, handcrafted signature, and clinical model.

As shown in Figure 2B, there were significant differences in the DLRN scores between the GR and PR groups in all cohorts (all p < 0.001). The DLRN showed a good performance for GR prediction in the TC with an AUC of 0.848 (95% CI, 0.794–0.901), which was further confirmed in all the validation cohorts, with AUCs larger than 0.800 (Figure 3). The stratified analyses revealed that the performance of DLRN was not affected by age, sex, BMI, tumor location, the version of CT, type of CT contrast agent, contrast agent concentration, contrast agent infused rate, and slice thickness (Fig. S5). Furthermore, DLRN showed significantly higher AUCs than the clinical model in all four cohorts, which also outperformed handcrafted signature in TC, IVC, and EVC2, as well as the DL signature in EVC1 and EVC1 (p < 0.05) (Table 2 and Fig. S6).

Figure 3

Receiver operating characteristic (ROC) curves of the four models. ROC curves of DLRN, deep learning signature, handcrafted signature, and clinical model, for predicting good responder (GR) in the (A) training cohort, (B) internal validation cohort, (C) external validation cohort 1, and (D) external validation cohort 2, respectively.

Table 2

Performance of models.

models	C-index (95%CI)
	training cohort	Internal Validation cohort	External Validation cohort 1	External Validation cohort 2	AIC
Handcrafted signature	0.693(0.617–0.769)	0.695(0.568–0.822)	0.737(0.649–0.825)	0.750(0.668–0.833)	698.91
DL signature	0.808(0.746–0.870)	0.806(0.705–0.907)	0.720(0.631–0.808)	0.734(0.642–0.827)	682.42
Clinical model	0.620(0.547–0.692)	0.518(0.404–0.633)	0.521(0.437–0.605)	0.626(0.551–0.702)	750.23
DLRN	0.848(0.794–0.901)	0.829(0.739–0.920)	0.804(0.732–0.877)	0.827(0.755–0.900)	585.65
	NRI (95%CI)				P values
DLRN vs Clinical	0.461(0.326–0.596)	0.473(0.264–0.682)	0.270(0.103–0.437)	0.508(0.356–0.660)	<0.001
	IDI (95%CI)				P values
DLRN vs Clinical	0.240(0.180–0.301)	0.245(0.152–0.339)	0.191(0.121–0.261)	0.247(0.168–0.326)	<0.001

Abbreviations: DL, deep learning; DLRS, deep learning radiomics signature; DLRS, deep learning radiomics nomogram.

The NRI and IDI analysis revealed that the integration of image signatures into the DLRN performed satisfactorily in all cohorts, indicating improved classification accuracy for GR prediction than the clinical model. Furthermore, the AIC values of the DLRN were lower than the corresponding clinical models or the two image signatures (Table 2). The calibration curves of the DLRN demonstrated that model-predicted GR was well-calibrated with the actual observation in all cohorts (p > 0.05) (Figure 2C). Additionally, the DCA graphically indicated that the DLRN provided a large net benefit than other models over the relevant threshold range in the whole cohorts (Figure 2D).

Preoperative predictors of survival

We also evaluated the prognostic value of DLRN in the 300 LAGC patients with follow-up data. The median follow-up period was 19 (range, 2–55) months. The LR, DM, or deaths from any cause occurred in 128 (42.7%) patients at a median follow-up period of 12 (range, 2–46) months.

Schoenfeld individual tests showed that the cox model fitted the proportional hazard assumption requirement (p > 0.05). We identified the optimal DLRN score to predict DFS as -0.313, and all patients were divided into high-risk and low-risk groups accordingly. The Kaplan–Meier curves demonstrated that higher DLRN scores were significantly associated with a better DFS (hazard ratio [HR], 0.886; 95% CI, 0.789–0.996, log-rank test, p = 0.003) (Figure 4A). Table 3 summarizes the results of univariable and multivariable Cox regression analysis of the predictors of DFS in the follow-up cohort, indicating that DLRN is an independent prognostic factor of DFS (HR, 0.500; 95% CI 0.283–0.886, p = 0.018), as well as diffuse tumors, poor differentiation and the cN3 stage before NCRT (Figure 4B). The final Cox regression model yielded a C-index of 0.693 (95% CI, 0.588–0.781).

Figure 4

Kaplan-Meier curves and forest plot of Disease-free survival (DFS) on the follow-up LAGC cohort. (A) Kaplan–Meier curves of DFS between the groups with low and high DLRN scores in the follow-up cohort. (B) Forest plot illustrating multivariable Cox regression analyses for DFS in the follow-up cohort.

Table 3

Uni- and multivariable cox regression analysis of predictors of disease-free survival.

	Univariable Analysis		Multivariable Analysis
Characteristics	Hazard ratio (95% CI)	P Value	Hazard ratio (95% CI)	P Value
Age	0.999(0.981–1.017)	0.924
BMI	0.971(0.921–1.022)	0.267
Sex (female vs. male)	0.784(0.518–1.185)	0.248
Tumor differentiation
Well and moderate	Ref		Ref
Poor	1.809(1.182–2.769)	0.006	1.740(1.129–2.680)	0.012*
Pre-CRT CEA (≤5 vs. >5)	1.235 (0.873–1.748)	0.234
Pre-CRT CA199(≤20 vs. >20)	1.385(0.939–2.044)	0.101
Location (%)
Cardia	Ref		Ref
Corpus	1.473(0.960–2.261)	0.076	1.132(0.731–1.753)	0.579
Antrum	1.170(0.736–1.861)	0.506	1.256(0.786–2.010)	0.340
Diffuse	2.768(1.507–5.083)	0.001	2.830(1.524–5.254)	0.001*
Clinical T stage
T2 and T3	Ref
T4a	1.298(0.888–1.897)	0.177
T4b	2.025(1.013–4.048)	0.046
Clinical N stage
0	Ref		Ref
1	1.070(0.544–2.105)	0.844	0.870(0.437–1.732)	0.692
2	1.196(0.620–2.306)	0.593	0.871(0.444–1.710)	0.688
3	2.893(1.544–5.422)	0.001	2.378(1.250–4.524)	0.008*
DLRN (low vs. high)	0.446(0.256–0.778)	0.004	0.500(0.283–0.886)	0.018*

Discussion

In the present study, we developed a non-invasive imaging signature by DL and radiomics analysis based on pretreatment venous-phase CT images, and independently validated its ability to predict response to NACT in large, multicenter cohorts. Furthermore, the imaging signature incorporated into the DLRN model exhibited improved performance in response prediction compared to the clinical models. Notably, the DLRN was significantly associated with DFS, providing useful complementary information about prognosis in patients with LAGC.

LAGC is clinically heterogeneous, necessitating accurate treatment response and prognosis prediction for the selection of appropriate treatment. Accumulating evidence has been reported to predict the response to NACT in patients with LAGC. Some texture features, such as entropy and delta gray-level cooccurrence matrix (GLCM) contrast, were found to be able to identify responders to NACT., In terms of radiomics, previous studies have focused on the application of CT-based features prior to NACT administration, displaying moderate performance with AUCs 0.70–0.75.18, 19, 20, Moreover, the detection radiomics model constructed by the baseline and restaging CT could be utilized for the early detection tasks of pathological downstaging (pDS), with excellent AUCs close to 0.90 in the internal validation cohort. Nevertheless, the optimistic result was mainly attributed to restage CT, which provided post-treatment tumor information and could not achieve an earlier prediction to guide NACT administration. However, these findings were of limited clinical relevance, due to the relatively small sample size and lack of validation in multicenter cohorts. Relatively, the lower AUC of 0.679 in the external validation cohort for the radiomics signature proposed in a larger population (323 cases) was more convincing, indicating that the radiomics features should be combined with auxiliary features to improve the prediction performance. Intriguingly, all the ten selected features used in our radiomics signature were transformation factors, especially the Laplacian of Gaussian (LoG) and wavelet-based features, providing more detailed information about tumoral heterogeneity.

A DL method, based on the DenseNet-121 architecture, was applied for DL feature extraction in this study. It was worth noting that, different from handcraft features, the DL method did not require delineated delineation, which not only reduces contour variability of different manual segmentation, but also enhances efficiency. Moreover, DL provides in-depth information included specific task in neural nets’ hidden layers without predefined features. The features captured by DL algorithm could predict lymph node metastasis, occult peritoneal metastasis, or survival outcomes for resectable GCs.,, The DL signature in our study presented a promising performance in GR prediction with AUCs larger than 0.72, similar to a previous study predicting the treatment response in esophageal squamous cell carcinoma by DL features. Similarly, the DL prediction model outperformed the handcrafted signature and the clinical model in discrimination ability in most of cohorts. All these indicated that DL offers a wealth of information reflecting tumor spatial heterogeneity and tumor microenvironment related to tumor chemosensitivity.

Furthermore, the prediction ability of DLRN was far better than that of the clinical model in all cohorts (p < 0.05). Previous studies have pointed out that various clinical or molecular risk factors are associated with neoadjuvant response. However, these metrics were not consistent across all studies. The distribution of cT, incorporated into our clinical model, was significantly different in some rather than all cohorts. Specifically, the AUC of the clinical model was only 0.518 in IVC, which was significantly lower than that of other models, including DLRN, handcrafted, and DL signatures. Tumor differentiation or no clinical factors were associated with treatment response in other studies., Additionally, the clinical factors only reflect specific aspects of the tumor. Patients with the same above features exhibited differential responses. This may be the reason for the ill-performance of the clinical model on different patient distribution. DLRN mines high-dimensional imaging features, followed by comprehensive quantification of intratumor heterogeneity, thereby improving the performance. In the present study, DLRN was constructed with 10 radiomics features, 18 DL features, and one clinical factor. These selected imaging features were not redundant but complementary, as shown in the heatmaps that these features are weakly correlated or uncorrelated. Moreover, the lowest AIC, and improved NRI and IDI also confirmed that the enhanced discrimination performance of DLRN was indeed due to feature integration instead of model overfitting.

Another interesting finding was that our DLRN is significantly associated with the DFS of patients with LAGC. Several studies proved that LAGC patients with different NACT treatment efficacy had varied prognosis., However, improved TRG after NACT might not necessarily mean prolonged survival. In the current study, compared to patients at high risk of GR, those at low risk of GR had greatly diminished the likelihood of favorable long-term outcomes even after NACT followed by curative resection. Patients with higher DLRN scores were significantly associated with a better DFS. In the multivariable Cox regression model, DLRN was an independent prognostic factor of DFS. Since the outcomes in low DLRN scores were not satisfactory, timely alternative curative-intent treatment approaches should be offered to avoid unnecessary toxicity and improve their survival outcomes. Accordingly, it is feasible to guide the treatment plan and implement personalized treatment based on our DLRN model.

Notably, the present study has some limitations. Firstly, although the multicenter sample size was larger, inherent bias was inevitable due to the retrospective nature and different distribution. Thus, further well-designed prospective studies are warranted to validate the generalizability and clinical applicability of our DLRN model. Secondly, we delineated the tumor on the largest slice with two-dimensional(2D) rather than the entire tumor(3D), which could not be representative of the entire tumor, and some radiomics features might be affected from 2D vs 3D. Hence, 3D analysis of the whole tumor deserved further investigation. Thirdly, for the survival analysis, the relative short-term follow-up part of a subset patients might increase the risk of type II error, with more censored data. A large number of patients in a multicenter study with prolonged follow-up time, and more unperceived clinicopathological factors that simultaneously impact the prognosis of LAGC after NACT, should be investigated further. Lastly, the biological significance of DLRN, especially DL features, still needs comprehensive elucidation. Thus, future investigations integrating imaging and molecular or gene data may provide insight into more micro information and their relationship.

In summary, we developed and validated a CT-based model using deep learning and radiomics analysis for the early prediction of response prior to NACT for patients with LAGC. The proposed DLRN, incorporating imaging signature and clinical factors, exhibited a promising performance for predicting response and clinical outcomes, and provide valuable information for individualized treatment in patients with LAGC. However, future prospective studies are required to confirm the clinical utility of our DLRN model.

Contributors

YC and XY were responsible for conception and design. YC, JZ, KW, LW, and ZS provided statistical analysis. All authors were involved in drafting and technical support in deep learning methods. JR assisted in statistical analysis. YC, JZ, and YL were involved in drafting the manuscript, and XM, XY, and XG were involved in reviewing the manuscript. YC and XY had full access to all the data. YC, ZL, XM, and XY verified the underlying data. All authors approved the final manuscript for submission.

Data sharing statement

Due to the privacy of patients, the data related to patients cannot be available for public access but can be obtained from the corresponding author on reasonable request approved by the institutional review board of all enrolled centers.

Declaration of interests

JR is an employee of GE Healthcare. YC received funding from National Natural Science Foundation of China (No. 82001789), China Postdoctoral Science Foundation (No. 2021M700897), and Project of Shanxi Provincial Health Commission (No. 2021XM51 and 2019058). ZL received funding from National Natural Science Foundation of China (No. 82001986), and Applied Basic Research Projects of Yunnan Province, China, Outstanding Youth Foundation (202101AW070001). YL received funding from National Natural Science Foundation of China (No. 82002702), and Youth Project of Natural Science Foundation of Hunan Science (No. 2020JJ5905). XY received funding from National Natural Science Foundation of China (No. 82171923), and Project of Shanxi Provincial Health Commission (No. 2020064). XG received funding from National Natural Science Foundation of China (No. 81871439). All other authors declare no competing interests.