Prediction model based on radiomics and clinical features for preoperative lymphovascular invasion in patients with gastric cancer with Lauren classification

doi:10.21203/rs.3.rs-2107626/v1

Download PDF

Research Article

Prediction model based on radiomics and clinical features for preoperative lymphovascular invasion in patients with gastric cancer with Lauren classification

https://doi.org/10.21203/rs.3.rs-2107626/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Objective

Lymphovascular invasion (LVI) is an independent risk factor of gastric cancer (GC) prognosis; however, LVI cannot be determined preoperatively. We explored whether a model based on contrast-enhanced computed tomography (CECT) radiomics features combined with clinicopathological factors can evaluate preoperative LVI in patients with GC with a clear Lauren classification.

Methods

We retrospectively analyzed 495 patients with GC, including 288 with LVI. The extracted CECT features were standardized, followed by consistency testing, correlation analysis, univariate analysis, and multivariate least absolute shrinkage and selection operator (LASSO) regression analysis. The radiomics score (Radcore) was calculated for each patient. Univariate (p < 0.10) and multivariate (p < 0.050) analyses were used to identify the clinical risk factors associated with LVI. Accordingly, three prediction models were established: Clinical + Arterial phase_Radcore, Clinical + Venous phase_Radcore, and a combined model. The prediction performances of the models were verified using receiver operating characteristics, decision curve analysis, and calibration curves in the testing dataset. The relationship between Lauren classification and LVI was analyzed using a histogram.

Results

The areas under the curve of the combined model were 0.8629 (95% confidence interval [CI], 0.8247–0.9011) and 0.8343 (95% CI, 0.7673–0.9012) in the training and testing datasets, respectively. The combined model had superior performance compared with the other models. Diffuse-type GC according to the Lauren classification accounted for 43.4% of LVI cases.

Conclusions

CECT-based radiomics models can effectively predict the preoperative LVI status in patients with GC with Lauren classification. The prediction ability of the models was effectively improved by incorporating clinicopathological factors.

Radiomics

Gastric cancer

LVI

Lauren classification

Gastric cancer (GC) is the fifth most common cancer and the third leading cause of cancer death worldwide [1]. In recent decades, it has ranked first among malignant tumors of the digestive tract in China [2], and more than 90% of cases are advanced GC [3]. Owing to the complicated and difficult treatment plan, the postoperative recurrence rate is as high as 40–50%, and the five-year survival rate remains unsatisfactory [2]. Contrast-enhanced computed tomography (CECT) is widely used in the preoperative examination of GC because it is convenient, fast, cost-effective, and minimally invasive [4]. Studies have shown that lymphovascular invasion (LVI) [5] is not only an important factor in tumor proliferation and metastasis [6] but also an independent predictor of GC prognosis. However, clinicians cannot accurately determine LVI status before surgery and can only judge LVI status through postoperative pathological examinations [7]. Therefore, a noninvasive method is urgently needed to help clinicians ascertain preoperative LVI status and promote personalized and accurate treatment [8].

Since 1965, the Lauren classification system has been widely accepted and used in clinical practice to classify GC into diffuse, intestinal, mixed, and undifferentiated types [9]. Studies have shown that vascular endothelial growth factor (VEGF) [10] synthesis and secretion are key factors in promoting heterogeneous angiogenesis and increasing vascular permeability [11, 12]. The VEGF values of diffuse type GC are higher than those of the intestinal type [11]. Moreover, angiogenesis increases the risk of LVI [3]. Therefore, there might a relationship between Lauren classification of GC and LVI.

Radiomics extracts a large number of imaging features from high-throughput imaging to reveal disease features that are indistinguishable to the naked eye and converts the data into quantitative features for tumor segmentation, feature extraction, and model building. Radiomics can improve the ability of physicians to make accurate diagnoses through deep mining, prediction, and the analysis of a large amount of image data [13]. Based on our understanding, radiomics studies focusing on GC are still in the exploratory phase [2]. Studies based on texture analysis of CT images have shown that CT radiomics features combined with clinicopathological factors such as T stage, N stage, and American Joint Committee on Cancer (AJCC) stage can be used to establish predictive models that can evaluate LVI status preoperatively [7]. Some studies used clinicopathological factors in combination with nomograms to predict vascular invasion in GC; however, these studies did not incorporate Lauren classification into their models as predictive factor, nor did they obtain quantitative information from CT images, such as the contrast enhancement rate [14]. Therefore, we established radiomics models based on CECT images to evaluate LVI status before surgery and proposed two hypotheses. First, incorporating clinicopathological factors such as N stage, T stage, cancer antigen (CA) 724, Lauren classification, and quantitative information obtained from CECT images into the models could improve their performance; second, Lauren classification of GC is associated with LVI.

Patients

This retrospective study was approved by the Institutional Review Board We retrospec,and waived the requirement for informed consent.tively analyzed all patients diagnosed with pathologically confirmed GC between January 2018 and December 2021. Finally, 495 cases were included in our study (age range: 33–84 years; mean age: 62.2 years old) (Table 1). The inclusion criteria were as follows: (1) GC was first confirmed on postoperative histopathology, (2) assessed by CECT before surgery, and (3) complete pathological diagnosis and clear Lauren classification. The exclusion criteria were as follows: (1) incomplete clinicopathological data, (2) poor CT image quality for any reason, (3) any local or systemic anticancer treatment before surgery, and (4) tumor thickness < 5 mm or well-differentiated tumor. We randomly divided the patients into training and testing datasets at a ratio of 7:3. The training dataset was used to build the models, and the testing dataset was used to validate the models built on the training dataset.

Table 1

Summary Descriptive Tables
Characteristics	Training dataset			Testing dataset			P*
Characteristics	0(N = 145)	1(N = 201)	p	0(N = 62)	1(N = 87)	p	P*
Gender			0.56			1.00	0.63
female	41(28.3%)	50(24.9%)		18(29.0%)	25(28.7%)
male	104(71.7%)	151(75.1%)		44(71.0%)	62(71.3%)
Age	60.0[53.0;68.0]	61.0[55.0;67.0]	0.66	60.5[55.2;64.0]	61.0[56.0;67.0]	0.26	0.81
Location			0.16			0.64	0.69
Cardia	2(1.38%)	12(5.97%)		1(1.61%)	4(4.60%)
Gastric antrum	74(51.0%)	103(51.2%)		32(51.6%)	43(49.4%)
Gastric body	5(3.45%)	7(3.48%)		2(3.23%)	2(2.30%)
Gastric fundus	5(3.45%)	5(2.49%)		5(8.06%)	3(3.45%)
Gastric horn	31(21.4%)	28(13.9%)		7(11.3%)	14(16.1%)
Large part of stomach	27(18.6%)	41(20.4%)		14(22.6%)	21(24.1%)
Pylorus	1(0.69%)	5(2.49%)		1(1.61%)	0(0.00%)
Thickness	15.0[11.0;20.0]	18.2[15.0;22.1]	< 0.001	15.0[11.0;20.0]	19.0[14.4;23.0]	0.010*	0.86
Borrmenn classification			0.14			0.35	0.76
I	6(4.14%)	7(3.48%)		3(4.84%)	1(1.15%)
II	31(21.4%)	27(13.4%)		14(22.6%)	12(13.8%)
III	96(66.2%)	146(72.6%)		41(66.1%)	66(75.9%)
IV	5(3.45%)	15(7.46%)		2(3.23%)	3(3.45%)
V	7(4.83%)	6(2.99%)		2(3.23%)	5(5.75%)
Pathological T stage			0.002*			0.42	0.27
T1	13(8.97%)	3(1.49%)		1(1.61%)	1(1.15%)
T2	25(17.2%)	25(12.4%)		13(21.0%)	13(14.9%)
T3	96(66.2%)	145(72.1%)		43(69.4%)	59(67.8%)
T4	11(7.59%)	28(13.9%)		5(8.06%)	14(16.1%)
Pathological N stage			< 0.001*			< 0.001*	0.87
N0	55(37.9%)	27(13.4%)		24(38.7%)	11(12.6%)
N1	26(17.9%)	29(14.4%)		15(24.2%)	13(14.9%)
N2	38(26.2%)	48(23.9%)		14(22.6%)	20(23.0%)
N3	26(17.9%)	97(48.3%)		9(14.5%)	43(49.4%)
Lauren classification			0.14			0.07	0.50
Diffuse type	50(34.5%)	89(44.3%)		16(25.8%)	36(41.4%)
Intestinal type	55(37.9%)	59(29.4%)		27(43.5%)	24(27.6%)
Mixed type	40(27.6%)	53(26.4%)		19(30.6%)	27(31.0%)
Differentiation status			0.003*			0.010*	0.83
High differentiation	11(7.59%)	4(1.99%)		4(6.45%)	2(2.30%)
Low differentiation	61(42.1%)	105(52.2%)		23(37.1%)	43(49.4%)
Medium and low differentiation	44(30.3%)	71(35.3%)		20(32.3%)	36(41.4%)
Medium differentiation	29(20.0%)	21(10.4%)		15(24.2%)	6(6.90%)
Infiltration style			0.17			0.69	0.15
a	38(26.2%)	36(17.9%)		15(24.2%)	16(18.4%)
b	49(33.8%)	72(35.8%)		26(41.9%)	39(44.8%)
c	58(40.0%)	93(46.3%)		21(33.9%)	32(36.8%)
CEA	2.20[1.32;3.89]	2.18[1.41;4.35]	0.28	2.33[1.48;3.74]	2.59[1.58;4.09]	0.48	0.26
CA724	9.42[6.10;20.2]	3.36[1.67;8.76]	< 0.001*	12.3[6.30;28.1]	2.96[1.38;6.31]	< 0.001*	0.97
CA199	2.02[1.31;6.63]	13.3[7.18;33.5]	< 0.001*	1.75[1.06;4.13]	13.3[6.00;25.9]	< 0.001*	0.05
Note:
0, non-lymphovascular invasionI; 1, lymphovascular invasion
The N in parentheses, the number of people
P* value, the significance of differences between the characteristics in training and testing cohorts
p value, the significance of differences between non-LVI group and LVI group
p < 0.05, statistical significance; P*< 0.05, statistical significance
CA724, Carbohydrate antigen 724
CA199, Carbohydrate antigen 199
CEA, Carcinoembryonic antigen

All patients underwent surgery within 2 weeks after CECT. Detailed clinical and pathological characteristics were collected from the electronic medical records of these patients (Table 1). Routine blood tests for CA724, CA199, and carcinoembryonic antigen (CEA) were performed 1 week before surgery. The normal levels of CA724, CA199, and CEA used in our hospital were 6.9 U/mL, 37.0 U/mL, and 5 ng/mL. TNM staging was reclassified according to AJCC/International Union for Cancer Control 8th edition. To reduce interobserver variability, pathology was histologically evaluated by two pathologists, specifically regarding Lauren classification and LVI status. Lesion thickness was measured simultaneously on the venous phase images by two radiologists, and the average of the two measurements was calculated.

The study workflow is presented in Fig. 1.

Image acquisition

All patients or legal guardians signed the consent form for CECT examination. Only one CT scanner (Somatom Definition AS, Siemens Healthineers, Cary, NC) was used for all patients in this study. Before scanning, all patients were required to fast for at least 8 h but to drink 600–1000 mL of water to achieve gastric distension. The use of anticholinergic drugs was not permitted. All patients were coached in breath-holding techniques before scanning. Before surgery, the participants underwent whole-abdominal or upper-abdominal CT using a 64-slice helical scanner in the supine position. Abdominal CT scan parameters were as follows: reconstructed slice thickness, 1 mm; reference tube current, 230mA; with automatic tube current modulation system; andtube voltage, 120 kV. A single-channel high-pressure syringe was connected to the patient’s venous indwelling needle and used to inject a non-ionic ioversol contrast medium (350 mg/mL, 3.5 mL/s). When the CT value of the descending aorta reached 100 HU, after a delay of 7 s, the arterial phase scan was performed. After the end of the arterial phase scan, the venous phase scan was performed after another 45-s delay.

Image segmentation

All venous- and arterial-phase CT images of the enrolled patients were acquired from the Picture Archive and Communication System (PACS) and uploaded to ITK-SNAP (http://www.itk-snap, version 3.6.4) for segmentation. Regions-of-interest (ROIs) were drawn manually along the tumor margins on all layers of the tumor on both the venous- and arterial-phase images, taking care to avoid artifacts and the gastric lumen. If multiple lesions of different sizes were found, only the largest lesion was selected for further analysis. Two radiologists with over 5 years of experience manually segmented the venous- and arterial-phase images of all patients. The radiologists were blinded to the clinical information. Before radiologist A performed tumor segmentation on all 495 patients, radiologists A and B performed simultaneous tumor segmentation on 30 patients randomly selected from the entire cohort, and radiologist A alone performed tumor segmentation on 30 patients. The above process ensured the consistency of lesion delineation. GC is characterized by gastric wall thickening or mass lesions, and CECT images are significantly enhanced. As reported in the literature [15], cancerous thickening was considered when focal thickening on axial images was ≥ 6 mm greater than the thickness of the adjacent gastric wall. If the location of the lesion was uncertain, the radiologist delineated the ROIs based on the pathology.

Radiomics features extraction and selection

We imported the volumes of interest (VOIs) generated from the ROIs of the venous- and arterial-phase images acquired with ITK-SNAP into AK software (Artificial Intelligence Suite V3.2.0, GE Healthcare, Chicago, IL) [16, 17] to extract features. To eliminate the per-feature unit limitation, we normalized the values of each feature in all images [18].

To verify the robustness of the radiomics features, we evaluated the feature extraction consistency between observers using the interclass correlation coefficient (ICC) [19]. ICC values were calculated using radiological features extracted from VOIs delineated by two radiologists. Any features with ICC < 0.9 were discarded. To remove redundant features, reduce model overfitting, and facilitate model interpretation, we used Spearman’s correlation analysis (SPM; the threshold for the correlation coefficient was 0.7) and univariate analysis (p < 0.050). Least absolute shrinkage and selection operator (LASSO) [20, 21] regression analysis was applied to the training dataset using the selected features to calculate each patient’s radiomics score (Radcore) (Arterial-phase Radcore [A_Rad], Venous-phase Radcore [V_Rad]).

Clinicopathologic characteristics selection

Clinical candidate variables included gender, age, location, thickness, Borrmann classification, T stage, tumor differentiation, N stage, Lauren classification, invasion pattern, CA724, CA199 and CEA. We first used univariate logistic regression to analyze clinical risk factors for LVI (P < 0.10), and then used multivariate logistic regression to identify clinical risk factors for LVI (p < 0.050).

Model building and validation

The forest plot [22] showed that A_Rad and V_Rad were statistically significant. Based on the clinical risk factors and each patient’s A_Rad and V_Rad, we established three models for LVI prediction: the Clinical + Arterial phase_Radcore (Cli + A_Rad), Clinical + Venous phase_Radcore (Cli + V_Rad), and combined models. We then developed a radiomics nomogram that included clinical and radiomics features [23, 24]. The radiomics nomogram provides clinicians with a quantitative tool to predict LVI status. The performances of the models were validated using a testing dataset. The receiver operating characteristic (ROC) curve could effectively measure the models’ ability to predict LVI status. Decision curve analysis (DCA) evaluated the clinical utility of the nomogram by quantifying the net benefit at different threshold probabilities. Calibration curves reflected the agreement between the predictive models and the observed outcomes.

Lauren classification

Histograms were used to reflect the proportions of various GC types according to the Lauren classification in the LVI and non-LVI groups. Meanwhile, considering the study of Wang et al. [11, 25–27], the relationship between the Lauren classification and LVI was briefly discussed.

Statistical analysis

In this retrospective analysis, we identified independent predictors associated with LVI status using univariate and multivariate logistic regression analyses. The ROC curve and area under the curve (AUC) were used to evaluate the predictive efficacy, sensitivity, and specificity of the models for LVI and non-LVI. Statistical significance was considered when p < 0.050. Data processing, calculations, radiomics nomograms, and forest plot construction were performed using R software (version 4.1.1).

Clinical and pathological characteristics

The clinical and pathological characteristics of the 495 patients are shown in Table 1. Among them, 288 (58.18%) patients had LVI. In the training dataset, T stage, N stage, differentiation status, lesion thickness, CA199, and CA724 were significantly different (all p < 0.050), whereas in the testing dataset, N stage, differentiation status, lesion thickness, CA199, and CA724 were significantly different (all p < 0.050). However, the clinical and pathological characteristics were not significantly different between the testing and training datasets (p > 0.050).

Radiomics feature selection

We extracted 1317 features from VOIs segmented from venous- and arterial-phase images. After we evaluated the selected features using ICCs, 1188 features were in the arterial phase, and 1182 features were in the venous phase. Subsequently, Spearman’s correlation analysis was used for feature selection, and 130 features were screened in the arterial phase and 120 in the venous phase. Using univariate analysis, 41 features in the arterial phase and 37 in the venous phase were identified. Finally, five features in the arterial phase and two features in the venous phase were screened using multivariate LASSO regression analysis. We then used the selected features to calculate each patient’s radiomics score (Radscore) as follows:

A_Rad = 0.3620 − 0.0539*original_shape_Elongation-0.3570*original_shape_Flatness + 0.3477*original_firstorder_90Percentile + 0.3588*original_glcm_Correlation + 0.4976*original_glcm_MaximumProbability

and

V_Rad = 0.3497 + 0.3003*original_firstorder_90Percentile-0.3858*original_glszm_SmallAreaEmphasis.

Clinical feature selection

According to the univariate analysis (p < 0.10), nine clinical risk factors were identified; then, according to the multivariate analysis (p < 0.050), five clinical risk factors showed statistical significance (Table 2). Although the Lauren classification was not significantly different between the LVI and non-LVI groups after univariate and multivariate logistic regression analyses, according to the literature [25–27], we believe that Lauren classification is an independent predictor of LVI. Thus, six characteristics were selected as independent predictors of LVI, i.e., Lauren classification, CA199, CA724, N stage, tumor location, and lesion thickness.

Table 2

Analysis of Clinicopathological Risk Factors Associated with LVI.
Characteristics	Ultivariate logistic regression			Multivariate logistic regression
Characteristics	OR	95%CI	P*	OR	95%CI	P**
Age	1.01	0.98–1.03	0.60	-	-	-
Borrmann classification
I	Reference	-	-	Reference	-	-
II	0.75	0.22–2.49	0.63	-	-	-
III	1.3	0.43-4	0.64	-	-	-
IV	2.57	0.58–11.38	0.21	-	-	-
V	0.73	0.16–3.43	0.70	-	-	-
CA199	1.02	1.01–1.03	< 0.001*	1.02	1.01–1.03	0.003*
CA724	0.99	0.98-1	0.03*	0.99	0.98-1	0.004*
CEA	1.01	1-1.02	0.14	-	-	-
Differentiation status
High differentiation	Reference	-	-	Reference	-	-
Low differentiation	4.73	1.44–15.51	0.01*	2.11	0.51–8.62	0.30
Medium and low differentiation	4.44	1.33–14.8	0.02*	1.58	0.41–6.05	0.51
Medium differentiation	1.99	0.56–7.12	0.29	0.74	0.18–2.98	0.67
Gender
female	Reference	-	-	Reference	-	-
male	1.19	0.73–1.93	0.48	-	-	-
Infiltration style
a	Reference	-		Reference	-	-
b	1.55	0.87–2.78	0.14	2.01	0.86–4.7	0.11
c	1.69	0.97–2.97	0.07*	1.07	0.42–2.71	0.89
Lauren classification
Diffuse type	Reference	-	-	Reference	-	-
Intestinal type	0.6	0.36-1	0.05*	1.22	0.44–3.37	0.70
Mixed type	0.74	0.44–1.27	0.28	0.58	0.25–1.32	0.20
Location
Cardia	Reference	-	-	Reference	-	-
Gastric antrum	0.23	0.05–1.07	0.06*	0.23	0.04–1.24	0.09
Gastric body	0.23	0.04–1.54	0.13	0.08	0.01–0.73	0.03
Gastric fundus	0.17	0.02–1.16	0.07*	0.17	0.02–1.5	0.11
Gastric horn	0.15	0.03–0.73	0.02*	0.16	0.03–0.94	0.04
Large part of stomach	0.25	0.05–1.22	0.09*	0.25	0.04–1.43	0.12
pylorus	0.83	0.06–11.42	0.89	0.64	0.04–11.23	0.76
Pathological N stage
N0	Reference	-	-	Reference	-	-
N1	2.27	1.13–4.58	0.02*	2.8	1.21–6.5	0.02*
N2	2.57	1.37–4.82	< 0.001*	2.66	1.25–5.66	0.01*
N3	7.6	4.04–14.3	< 0.001*	6.5	2.95–14.32	< 0.001*
Pathological T stage
T1	Reference	-	-	Reference	-	-
T2	4.33	1.1-17.09	0.04*	3.46	0.78–15.41	0.10
T3	6.55	1.82–23.58	< 0.001*	2.05	0.78–15.42	0.32
T4	11.03	2.62–46.38	< 0.001*	2.52	0.78–15.43	0.27
Thickness	1.08	1.04–1.11	< 0.001*	1.06	0.78–15.44	0.005*
Note:
p* < 0.1, statistical significance
p**< 0.05, statistical significance
CA724, Carbohydrate antigen 724
CA199, Carbohydrate antigen 199
CEA, Carcinoembryonic antigen
LVI, lymphovascular invasion.

Model building and validation

According to the forest plot (Fig. 2), A_Rad and V_Rad were statistically significant (p < 0.050). We built three prediction models based on A_Rad, V_Rad, and the six predictors and created a radiomics nomogram (Fig. 3).

The ROC curve (Fig. 4) shows the discriminative performance of each predictive model. In the training dataset, the AUC of the combined model was 0.8629 (95% confidence interval [CI], 0.8274–0.9011), the specificity was 0.703, and the sensitivity was 0.851 (Fig. 4e). In the testing dataset, the AUC of the combined model was 0.8343 (95% CI, 0.7673–0.9012), the specificity was 0.828, and the sensitivity was 0.774 (Fig. 4f). The combined model showed better performance than the Cli + A_Rad and Cli + V_Rad models. The calibration curve demonstrated good agreement between the training (p = 0.33) (Fig. 4a) and testing datasets (p = 0.57) (Fig. 4b). DCA of both the training dataset (Fig. 4c) and the testing dataset (Fig. 4d) showed that the combined model had a better cost-to-benefit ratio at the most reasonable threshold probability than the other models.

Lauren classification

In this retrospective study, 288 patients had LVI, of which 125 (43.4%) had diffuse type, and 83 (28.8%) had intestinal type. Among the 207 patients in the non-LVI group, 66 (31.9%) had diffuse type, and 82 (39.6%) had intestinal type (Fig. 5).

CECT is a common preoperative examination for GC and is widely used in clinical practice [28, 29]. In this retrospective study, we transformed CT image features into quantitative features and developed and validated models based on a combination of radiomics and clinical features for preoperative individualized noninvasive prediction of GC LVI status. To the best of our knowledge, the Lauren classification and quantitative features were incorporated into the models to assist with predicting the LVI status for the first time. Previous studies [14, 30–33] have shown that Lauren classification and LVI status are independent factors affecting the efficacy of and prognosis after chemotherapy, indicating that our research has important significance in the treatment and prognosis of GC.

Univariate and multivariate analyses revealed that T stage, thickness, location, CA199, and CA724 were statistically significant predictors of preoperative LVI status (p < 0.050). Chen et al. [7] found that CA199 was not a significant predictive factor, which was inconsistent with our findings and may be owing to our larger sample size and more detailed clinical information. For the CECT images, we used the thickness of the tumor to represent the depth of tumor invasion. Li et al. [34] retrospectively analyzed the pathological data of 361 non-metastatic GC patients and found that the LVI detection rate was 13.9%. This was significantly related to depth of invasion, TNM stage, and tumor size, which was consistent with our findings.

The main focus of this study was to construct LVI prediction models integrating radiomics features, Lauren classification, quantitative features, CA199 and CA724, tumor location, and N stage. The Lauren classification is the most widely used classification system for GC [2]. Some studies have shown that Lauren classification is related to LVI [3, 11]; accordingly, we included Lauren classification as an independent predictor in our prediction models. Meng et al. [14] used clinicopathological information such as the degree of differentiation and preoperative pathological stage and constructed a nomogram to evaluate LVI status, with AUCs of 0.792 and 0.774 for the training and testing datasets, respectively. Chen et al. [7] used N staging, T staging, and AJCC staging combined with radiological features to establish predictive models to evaluate the LVI status of patients with GC. The AUCs of their training and testing datasets were 0.856 and 0.792, respectively. We constructed models that combined N staging, lesion thickness, Lauren classification, tumor location, CA199, CA724, and Radscores. The combined model demonstrated strong predictive power with an AUC of 0.8629, sensitivity of 0.851, and specificity of 0.703 for the training dataset and an AUC of 0.8343, sensitivity of 0.774, and specificity of 0.828 for the testing dataset. The predictive ability of the model was improved compared with the previously reported models.

In this study, 288 patients developed LVI, including 125 with diffuse type accounting for 43.4% of the total number of patients with LVI. Diffuse GC according to the Lauren classification accounted for the highest proportion of LVI patients, which is consistent with the findings of Li et al. [27].

The strengths of our study are as follows. First, the clinical and pathological data of the patients included in this study were complete. We were concerned with the influence of the Lauren classification on LVI and incorporated it into the prediction models, which was more instructive for clinical treatment. Second, quantitative features were added to the study. In the CECT images, we used the thickness of the lesion to represent the invasion depth. In the CECT venous-phase images, two physicians reassessed the thickness of the lesions. Although this assessment method has a measurement error, including lesion thickness is necessary because a study [3] has shown that the depth of invasion is one of the risk factors for LVI in patients with GC. Third, we normalized the extracted radiomics features to address the interpretability challenge of radiomics features and calculated the value of radiomics features to convenient for clinical work. In addition, we used different feature selection methods to extract features to ensure better stability and accuracy of the extracted features. Fourth, although the images in this study were manually segmented, we used the ICC to evaluate the agreement of image segmentation between the two radiologists, thus ensuring the reliability of the extracted features. Based on the above discussion, we believe that the proposed model is reliable and stable and can help radiologists make a correct diagnosis. Fifth, we quantified the predictive models using a visualization and interpretability tool, namely the radiomics nomogram. The radiomics nomogram is an easy-to-use personalized decision-making tool that assists clinicians in understanding the LVI status before surgery and facilitates personalized precision medicine.

This study had several limitations. First, few quantitative features were extracted from the CT images; only lesion thickness was incorporated into the models, and the models were mostly based on clinicopathological factors. Therefore, our analysis may have been subject to selection bias. Second, this was a single-center study, and prospective multi-center experimental studies are still needed to experimentally validate these models. Third, subjective factors may exist in the Lauren classification, leading to classification errors. In this study, the relationship between the Lauren classification and LVI was limited to the most basic discussion, and other methods will be used for further exploration.

A radiomics model consisting of CECT radiomics features combined with clinicopathological factors can effectively predict preoperative LVI status in patients with GC with a Lauren classification. We integrated the Lauren classification, quantitative features, N stage, tumor location, and CA199 and CA724 levels to effectively improve the prediction ability of the models.

Data availability

The corresponding author may provide the data involved in this study upon reasonable request

Conflicts of interest

We have no conflict of interest

Funding

This study was funded by National Natural Science Foundation of China (grant number 81901701) and the Postdoctoral Science Foundation of the Ministry of Heilongjiang Province (grant number LBH-Q20123)

Acknowledgments

The authors thank all those in the Radiology Department of the Cancer Hospital affiliated to Harbin Medical University who contributed to this article. Thanks to Editage (www.editage.cn) for polishing the English of this article.

Conflict of Interest Statement

The authors of this article declare that there is no potential conflict of interest with any organization or individual.

Ethical approval

The study was reviewed and approved by the Ethics Committee of the Cancer Hospital Affiliated to Harbin Medical University and was in accordance with the Declaration of Helsinki (revised in 2013), and waived the requirement for informed consent.

AUTHOR CONTRIBUTIONS

PW and KC create datasets, interpret data, and define clinical labels. YH, AN, JS, SY, and HP collected patient clinical data and collated patient information. MZ provided method support and statistical analysis. PW wrote the manuscript. WM and NS provided clinical expertise and guidance for study design and oversaw the project. All authors contributed to the article and approved the submitted version.

Rong D, Lu C, Zhang B, Fu K, Zhao S, Tang W, et al. CircPSMC3 suppresses the proliferation and metastasis of gastric cancer by acting as a competitive endogenous RNA through sponging miR-296-5p. Mol Cancer. 2019;18:25.
Sun Z, Jin L, Zhang S, Duan S, Xing W, Hu S. Preoperative prediction for lauren type of gastric cancer: A radiomics nomogram analysis based on CT images and clinical features. J Xray Sci Technol. 2021;29:675–86.
Li F, Chen Z, Tan B, Liu Y, Zhao Q, Fan L, et al. Influential factors and prognostic analysis of blood vessel invasion in advanced gastric cancer. Pathol Res Pract. 2020;216:152727.
Naha PC, Hsu JC, Kim J, Shah S, Bouche M, Si-Mohamed S, et al. Dextran-Coated Cerium Oxide Nanoparticles: A Computed Tomography Contrast Agent for Imaging the Gastrointestinal Tract and Inflammatory Bowel Disease. ACS Nano. 2020;14:10187–97.
Ma Z, Liang C, Huang Y, He L, Liang C, Chen X, et al. Can lymphovascular invasion be predicted by preoperative multiphasic dynamic CT in patients with advanced gastric cancer? Eur Radiol. 2017;27:3383–91.
Liu S, Liu S, Ji C, Zheng H, Pan X, Zhang Y, et al. Application of CT texture analysis in predicting histopathological characteristics of gastric cancers. Eur Radiol. 2017;27:4951–9.
Chen X, Yang Z, Yang J, Liao Y, Pang P, Fan W, et al. Radiomics analysis of contrast-enhanced CT predicts lymphovascular invasion and disease outcome in gastric cancer: a preliminary study. Cancer Imaging. 2020;20:24.
Ma J, Shen H, Kapesa L, Zeng S. Lauren classification and individualized chemotherapy in gastric cancer. Oncol Lett. 2016;11:2959–64.
Qiu M, Zhou Y, Zhang X, Wang Z, Wang F, Shao J, et al. Lauren classification combined with HER2 status is a better prognostic factor in Chinese gastric cancer patients. BMC Cancer. 2014;14:823.
Leung DW, Cachianes G, Kuang WJ, Goeddel DV, Ferrara N. Vascular endothelial growth factor is a secreted angiogenic mitogen. Sci (New York NY). 1989;246:1306–9.
Ma L, Xu X, Zhang M, Zheng S, Zhang B, Zhang W, et al. Dynamic contrast-enhanced MRI of gastric cancer: Correlations of the pharmacokinetic parameters with histological type, Lauren classification, and angiogenesis. Magn Reson Imaging. 2017;37:27–32.
Li Z, Zhao Z, Wang C, Wang D, Mao H, Liu F, et al. Association Between DCE-MRI Perfusion Histogram Parameters and EGFR and VEGF Expressions in Different Lauren Classifications of Advanced Gastric Cancer. Pathol Oncol Res. 2021;27:1610001.
Lambin P, Rios-Velazquez E, Leijenaar R, Carvalho S, van Stiphout RG, Granton P, et al. Radiomics: extracting more information from medical images using advanced feature analysis. Eur J Cancer. 2012;48:441–6.
Meng Y, Huang X, Liu J, Chen J, Bu Z, Wu G, et al. A Novel Nomogram for Individually Predicting of Vascular Invasion in Gastric Cancer. Technol Cancer Res Treat. 2021;20:15330338211004924.
Wang N, Wang X, Li W, Ye H, Bai H, Wu J, et al. Contrast-Enhanced CT Parameters of Gastric Adenocarcinoma: Can Radiomic Features Be Surrogate Biomarkers for HER2 Over-Expression Status? Cancer Manag Res. 2020;12:1211–9.
Shu Z, Fang S, Ye Q, Mao D, Cao H, Pang P, et al. Prediction of efficacy of neoadjuvant chemoradiotherapy for rectal cancer: the value of texture analysis of magnetic resonance images. Abdom Radiol (NY). 2019;44:3775–84.
Gao C, Xiang P, Ye J, Pang P, Wang S, Xu M. Can texture features improve the differentiation of infiltrative lung adenocarcinoma appearing as ground glass nodules in contrast-enhanced CT? Eur J Radiol. 2019;117:126–31.
Zwanenburg A. Radiomics in nuclear medicine: robustness, reproducibility, standardization, and how to avoid data analysis traps and replication crisis. Eur J Nucl Med Mol Imaging. 2019;46:2638–55.
Ionan AC, Polley M-YC, McShane LM, Dobbin KK. Comparison of confidence interval methods for an intra-class correlation coefficient (ICC). BMC Med Res Methodol. 2014;14:121.
Hu P, Liu Y, Li Y, Guo G, Su Z, Gao X, et al. A Comparison of LASSO Regression and Tree-Based Models for Delayed Cerebral Ischemia in Elderly Patients With Subarachnoid Hemorrhage. Front Neurol. 2022;13:791547.
Kang J, Choi YJ, Kim IK, Lee HS, Kim H, Baik SH, et al. LASSO-Based Machine Learning Algorithm for Prediction of Lymph Node Metastasis in T1 Colorectal Cancer. Cancer Res Treat. 2021;53:773–83.
Andrade C. (2020) Understanding the Basics of Meta-Analysis and How to Read a Forest Plot: As Simple as It Gets. J Clin Psychiatry 81.
Tan Y, Zhang S-T, Wei J-W, Dong D, Wang X-C, Yang G-Q, et al. A radiomics nomogram may improve the prediction of IDH genotype for astrocytoma before surgery. Eur Radiol. 2019;29:3325–37.
Zhang W, Mao N, Wang Y, Xie H, Duan S, Zhang X, et al. A Radiomics nomogram for predicting bone metastasis in newly diagnosed prostate cancer patients. Eur J Radiol. 2020;128:109020.
Wang XX, Ding Y, Wang SW, Dong D, Li HL, Chen J, et al. Intratumoral and peritumoral radiomics analysis for preoperative Lauren classification in gastric cancer. Cancer Imaging. 2020;20:83.
Park DJ, Yoon C, Thomas N, Ku GY, Janjigian YY, Kelsen DP, et al. Prognostic significance of targetable angiogenic and growth factors in patients undergoing resection for gastric and gastroesophageal junction cancers. Ann Surg Oncol. 2014;21:1130–7.
Li Y, Xue XW, Luo YF, Wu HW, Chen J, Zhou WX. [Clinicopathologic features of gastric adenocarcinoma based on the revised Lauren's classification]. Zhonghua Bing Li Xue Za Zhi = Chinese Journal of Pathology. 2018;47:486–91.
Li W, Zhang L, Tian C, Song H, Fang M, Hu C, et al. Prognostic value of computed tomography radiomics features in patients with gastric cancer following curative resection. Eur Radiol. 2019;29:3079–89.
Liang P, Ren XC, Gao JB, Chen KS, Xu X. Iodine Concentration in Spectral CT: Assessment of Prognostic Determinants in Patients With Gastric Adenocarcinoma. AJR Am J Roentgenol. 2017;209:1033–8.
Qiu M-z, Cai M-y, Zhang D-s, Wang Z-q, Wang D-s, Li Y-h, et al. Clinicopathological characteristics and prognostic analysis of Lauren classification in gastric adenocarcinoma in China. J Translational Med. 2013;11:58.
Tang CT, Zeng L, Yang J, Zeng C, Chen Y. Analysis of the Incidence and Survival of Gastric Cancer Based on the Lauren Classification: A Large Population-Based Study Using SEER. Front Oncol. 2020;10:1212.
Wu Y-f, Xu H-m, Chen J-q. [Evaluation of the correlations of cell proliferation activity with lymphatic and vascular invasion and prognosis in gastric carcinoma]. Zhonghua zhong liu za zhi [Chinese journal of oncology]. 2005;27:492–5.
Meng Y, Huang X, Liu J, Chen J, Bu Z, Wu G, et al. A Novel Nomogram for Individually Predicting of Vascular Invasion in Gastric Cancer. Technol Cancer Res Treat. 2021;20:15330338211004924.
Li P, Ling Y-H, Zhu C-M, Hu W-M, Zhang X-K, Luo R-Z, et al. Vascular invasion as an independent predictor of poor prognosis in nonmetastatic gastric cancer after curative resection. Int J Clin Exp Pathol. 2015;8:3910–8.

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Prediction model based on radiomics and clinical features for preoperative lymphovascular invasion in patients with gastric cancer with Lauren classification

Status:

Version 1

Abstract

Objective

Methods

Results

Conclusions

Figures

Introduction

Materials And Methods

Patients

Image acquisition

Image segmentation

Radiomics features extraction and selection

Clinicopathologic characteristics selection

Model building and validation

Lauren classification

Statistical analysis

Results

Clinical and pathological characteristics

Model building and validation

Lauren classification

Discussion

Conclusion

Declarations

References

Additional Declarations

Status:

Version 1