The value of CT radiomics features to predict visceral pleural invasion in ≤3 cm peripheral type early non-small cell lung cancer

Abstract

OBJECTIVE:

To investigate predictive value of CT-based radiomics features on visceral pleural invasion (VPI) in ≤3.0 cm peripheral type early non-small cell lung cancer (NSCLC).

METHODS:

A total of 221 NSCLC cases were collected. Among them, 115 are VPI-positive and 106 are VPI-negative. Using a stratified random sampling method, 70% cases were assigned to training dataset (n = 155) and 30% cases (n = 66) were assigned to validation dataset. First, CT findings, imaging features, clinical data and pathological findings were retrospectively analyzed, the size, location and density characteristics of nodules and lymph node status, the relationship between lesions and pleura (RAP) were assessed, and their mean CT value and the shortest distance between lesions and pleura (DLP) were measured. Next, the minimum redundancy-maximum relevance (mRMR) and least absolute shrinkage and selection operator (LASSO) features were extracted from the imaging features. Then, CT imaging prediction model, texture feature prediction model and joint prediction model were built using multifactorial logistic regression analysis method, and the area under the ROC curve (AUC) was applied to evaluate model performance in predicting VPI.

RESULTS:

Mean diameter, density, fractal relationship with pleura, and presence of lymph node metastasis were all independent predictors of VPI. When applying to the validation dataset, the CT imaging model, texture feature model, and joint prediction model yielded AUC = 0.882, 0.824 and 0.894, respectively, indicating that AUC of the joint prediction model was the highest (p < 0.05).

CONCLUSION:

The study demonstrates that the joint prediction model containing CT morphological features and texture features enables to predict the presence of VPI in early NSCLC preoperatively at the highest level.

Keywords

Non-small cell lung cancer (NSCLC)Visceral pleural invasion (VPI)Computed tomography (CT)radiomics predictive models

1 Introduction

Lung cancer is one of the leading causes of cancer death worldwide, among which non-small cell lung cancer (NSCLC) accounts for about 85% of the cases, with a 5-year overall survival rate (OS) of only 21% [1]. Visceral pleural invasion (VPI) is one of the most important factors for poor prognosis of lung cancer and an important predictor of postoperative recurrence and lymph node metastasis [2, 3]. In the eighth edition of the TNM staging criteria of the International Association for the Study of Lung Cancer (IASLC), it is recommended to upgrade the T1 stage to T2 stage and stage Ia to stage Ib for NSCLC ≤3.0 cm with VPI [4]. While tumor TNM stage will directly affect the surgical resection method and the need for adjuvant chemotherapy after surgery [5], which affects patient prognosis.

Currently, the diagnosis of lung cancer VPI relies on histopathology as the gold standard, and CT is the preferred tool for preoperative prediction of lung cancer VPI. The manifestations such as the length of the contact surface between the tumor and the pleura >3.0 cm, the obtuse angle connected to the pleura, and the significant thickening of the pleura, and the presence of ≥2 signs can be used to diagnose VPI, however, the sensitivity and specificity of these methods are not high [6].

Radiomics can extract non-visual image information features to assess the intrinsic heterogeneity and aggressiveness of tumors in high throughput, and has been widely used in the differentiation of benign and malignant lung nodules [7], prediction of gene mutations [8], lung cancer staging [9] and prognosis assessment [10]. However, radiomics studies of VPI are poorly reported in the literature, especially for early-stage lung cancer. In this study, the correlation between CT performance combined with quantitative radiomics features and VPI was retrospectively analyzed. The aim of this study was to explore the value of radiomics combined with CT features on predicting VPI in ≤3.0 cm peripheral type early NSCLC.

2 Information and methods

2.1 General study data

The data of 221 patients with surgically pathologically confirmed ≤3.0 cm peripheral type NSCLC between October 2015 and October 2021 in Anhui Cancer Hospital were collected.

Inclusion criteria: (1) maximum diameter of lesion ≤3.0 cm, (2) pathologically confirmed peripheral NSCLC, (3) complete preoperative image archiving and communication system (PACS) data, (4) interval between CT scan and surgery ≤2 weeks; (5) lung segment resection only.

Exclusion criteria: (1) lesion >3.0 cm; (2) distant metastasis before surgery; (3) puncture biopsy before CT scan and history of radiotherapy; (4) incomplete PACS data. This retrospective study was agreed and approved by the hospital ethics committee to waive informed consent for all patients (Approval number: 2022-YXK-04).

In all collected cases, 70% of these patients were randomly assigned to the training set (n = 155) to establish the predictive models and the remaining patients (n = 66) were assigned to the validation set to evaluate the performance of the models using a stratified random sampling method.

2.2 CT scanning protocol

Transverse scans were performed using GE Discovery64-slice CT in the supine position, ranging from the lower neck scan to the adrenal level. Scanning parameters: tube voltage 120 kVp, adaptive tube current, layer thickness 1.25 mm, pitch 0.984:1, rotation time 0.5 s, FOV 350 mm×350 mm, matrix 512×512, layer thickness 5.0 mm, spacing 5.0 mm; lung window settings: window width 1500 HU, window level-500 HU; Mediastinal window setting: window width 400 HU, window level 40 HU; non-ionic contrast medium iohexol (1.5 mL/kg, 350 mg I/mL) was injected at 2.5 mL/s via the cubital vein on contrast-enhanced scans; 1.25 mm slice thickness reconstruction was performed using the 40% ASiR iterative reconstruction technique after the end of the scan.

2.3 Evaluation of CT images

All images were analyzed on GE ADW4.6 workstation by two diagnostic thoracic imaging physicians with more than 10 years of experience using the multiple planar reconstruction (MPR) technique, and in case of disagreement on the analysis and measurement results, consensus was reached by mutual agreement. The nodule location, size and density, and relationship of the lesion to adjacent pleura (RAP) were evaluated, mean nodule diameter, CT mean and minimum distance from the lesion to the pleura (DLP) signs were measured.

RAP was classified into 6 types [11] (Fig. 1A–F): Type I “linear sign”: one or more linear splines extended from the tumor surface to the pleura without pleural reaction; Type II “natural pleural sign”: the tumor was close to the pleural surface and the pleura was naturally shaped without thickening. Type III “pleural thickening sign”: the tumor was close to the pleural surface, and the pleura was thickened with shrinkage; Type IV “trumpet sign”: one or more linear splines extended from the tumor surface to the pleura with triangular soft tissue density shadows at the pleural end; Type V type V “triangular sign”: the pleura showed triangular soft tissue density shadows and the tip abuts the tumor surface; type VI “umbilical concave sign”: the pleura was drawn into the tumor in an arc-shaped change; types IV, V and VI were marked as pleural indentation PI; when multiple forms existed, they were preferentially labeled as VI, V, IV, III, II and I according to the weight.

Fig. 1

(A) Type I “linear sign”: male, poorly differentiated adenocarcinoma of the right upper lung, VPI (–). (B) Type II “natural edge sign”: highly differentiated adenocarcinoma of the left lower lung, VPI (–). (C) Type III “thickening edge sign”: middle differentiated adenocarcinoma of the left lower lung, VPI (+). (D) Type IV “trumpet sign”: highly differentiated adenocarcinoma at the tip of the left upper lung, VPI (+). (E) Type V “triangle sign”: highly differentiated adenocarcinoma of the left upper lung, VPI (+). (F) Type VI “umbilical concave sign”: highly differentiated adenocarcinoma of right upper lung, VPI (+). (G) The tumor cells were located under the pleura and did not involve elastic fiber (EFs×400) PL0. (H) Tumor cells invade the elastic fibrous layer with interruption (EFs×400) PL1. (I) Lung window. (J) Magnification and segmentation of nodules.

2.4 Texture feature extraction

Using ITK-SNAP software, the region of interest (ROI, Fig. 1I∼J) was manually outlined along the nodal edge at the largest level of the lesion in the high-resolution CT scan lung window, and the original images and ROI were imported into the A.K. (Artificial Intelligence Kitl Version V3.2.2.R) post-processing software provided by GE, USA, and after voxel adjustment, resampling, noise reduction and grayscale discretization, normalized reconstruction and fusion were performed. For each ROI, firstorder features, Gray Level Co-occurrence Matrix (GLCM) features, Gray Level Size Zone Matrix (GLSZM) features, and Gray Level Run Length Matrix (GLRL) features were extracted.

2.5 Reproducibility assessment

Reproducibility assessment was performed by two senior chest imaging experienced physicians (A and B) with unknown clinicopathological findings. Physician A completed CT sign evaluation, ROI outline and feature extraction for all data twice, which were used to evaluate the measurer’s own consistency; physician B only performed CT sign evaluation, ROI setting and feature extraction once, which was used to evaluate the consistency with physician A. Inter-group and intra-group consistency coefficients ICC >0.75 indicated good consistency.

2.6 Texture feature analysis and model construction

The IPMs software was used to reduce the dimensionality of all texture parameters; two feature selection methods, mRMR and LASSO, were selected for the descending dimension method. Firstly, mRMR was performed to remove redundant and irrelevant features; then LASSO was performed (Fig. 2) to select the optimal feature subset; and the corresponding model was constructed to evaluate the diagnostic effectiveness by ROC curve.

Fig. 2

Feature Selection. (A∼B) The LASSO includes choosing the regular parameter λ and determining the number of the feature. (C) After the number of features determined, the most predictive subset of feature is chosen and the corresponding coefficients are evaluated.

2.7 Pathological analysis

Two diagnostic thoracic pathologists with 10 years of experience assessed VPI by blind methods. When hematoxylin-eosin (HE) stained slides showed lesions adjacent to the pleura and the presence or absence of pleural invasion could not be clearly distinguished, elastic fiber (EFs) staining was performed to determine the presence of VPI. According to the pleural invasion grading classification of the IASLC 8th edition [4] TNM staging scheme for lung cancer: PL0 refers to no invasion of the visceral pleura (Fig. 1G), PL1 refers to tumor invasion of the elastic fiber layer of the visceral pleura (Fig. 1H), PL2 refers to tumor breaking through the surface mesothelial layer of the visceral pleura or falling into the chest cavity, and PL3 refers to invasion of the wall pleura or chest wall. In this study, PL1 and PL2 were categorized as VPI (+) group with tumor diameter ≤3.0 cm and VPI (+) as T2a stage, and as stage Ib due to the absence of lymph nodes or distant metastasis; while PL0 was categorized as VPI (–) group as T1 stage, and as stage Ia due to the absence of lymph nodes or distant metastasis.

2.8 Statistical analysis

K-S test (test level P = 0.05) was used for normality test of measurement data, Levene method (test level P = 0.05) was used for homogeneity test of variance, data in accordance with normal distribution and homogeneity of variance were expressed as ‘mean±standard deviation’, and t-test was used for comparison of measurement data. Data not conforming to normal distribution or homogeneity of variance were expressed as median, and the Mann Whitney U test was used for comparison between groups. Count data were expressed as frequencies and percentages. Two groups were compared using the chi-square test or Fisher exact test according to theoretical frequencies. Logistic regression analysis was used to construct prediction models, and their diagnostic efficacies were assessed by ROC curves.

All statistics were tested using a two-sided test at the level of a = 0.05. P < 0.05 indicated that the differences were statistically significant. SPSS (Version 20), MedCalc (Version 16.8.4) and IPMs (Version 1.1.1.R) software were applied to organize and analyze the collected data, and the stratified random sampling method was used to randomly divide the data into training set and validation set (7:3).

3 Results

3.1 The demographic information of patients enrolled

A total of 221 patients, aged 38–86 years, with an average age of 61.28 years, 97 males and 124 females, were included in the study, and 115 patients were VPI positive and 106 were VPI negative according to the pathology. The differences of demographic information such as age, sex, location and disease typing between VPI positive and VPI negative patients in training set or validation set were not statistically significant (P > 0.05, Table 1).

Table 1
Results of comparative analysis of CT imaging signs

Signs Category Training Set (n = 155) P Validation Set (n = 66) P

VPI-positive VPI-negative VPI-positive VPI-negative

(n = 81) (n = 74) (n = 34) (n = 32)

Age Mean value (years) 60.42±10.13 61.38±10.18 0.558 62.09±10.18 62.50±11.83 0.880

Gender Male 32 (39.51%) 36 (48.65%) 0.262 16 (47.06%) 14 (43.75%) 0.810

Female 49 (60.49%) 38 (51.35%) 18 (52.94%) 18 (56.25%)

Size Mean diameter (mm) 19.91±5.95 16.88±5.58 0.001 18.95±7.17 15.82±4.52 0.039

CT value Mean value (HU) 28.10 (26.10) –62.65 (268.98) <0.001 32.05 (18.68) –26.80 (234.83) 0.001

Density Pure ground glass 0 (0.00%) 3 (4.05%) <0.001 1 (2.94%) 0 (0.00%) <0.001

Hybrid ground glass 19 (23.46%) 40 (54.05%) 3 (8.82%) 16 (50.00%)

Solidity 62 (76.54%) 31 (41.90%) 30 (88.24%) 16 (50.00%)

Department (right middle lobe) (counted as upper lung) Upper right lung 35 (43.21%) 35 (47.29%) 0.888 15 (44.12%) 15 (46.88%) 0.498

Lower right lung 13 (16.05%) 13 (17.57%) 9 (26.47%) 6 (18.75%)

Upper left lung 24 (29.63%) 20 (27.03%) 9 (26.47%) 7 (21.87%)

Lower left lung 9 (11.11%) 6 (8.11%) 1 (2.94%) 4 (12.50%)

Shortest distance between lesion and pleura DLP (mm) 0.01 (0.00) 1.00 (3.00) <0.001 0.00 (0.25) 2.00 (4.83) 0.004

Relationship with pleura Type I 6 (7.41%) 18 (24.32%) <0.001 1 (2.94%) 9 (28.12%) 0.002

Type II 2 (2.47%) 12 (16.22%) 1 (2.94%) 4 (12.50%)

Type III 8 (9.88%) 8 (10.82%) 2 (5.88%) 4 (12.50%)

Type IV 14 (17.28%) 16 (21.62%) 8 (23.53%) 9 (28.12%)

Type V 23 (28.39%) 10 (13.51%) 13 (38.24%) 3 (9.38%)

Type VI 28 (34.57%) 10 (13.51%) 9 (26.47%) 3 (9.38%)

Pleural depression sign There were 65 (80.25%) 36 (48.65%) <0.001 30 (88.24%) 15 (46.87%) <0.001

None 16 (19.75%) 38 (51.35%) 4 (11.76%) 17 (53.13%)

Degree of differentiation High 49 (60.50%) 58 (78.38%) 0.041 14 (41.18%) 22 (68.75%)

Medium 18 (22.22%) 7 (9.46%) 10 (29.41%) 9 (28.12%)

Low 14 (17.28%) 9 (12.16%) 10 (29.41%) 1 (3.13%)

Pathological typing Adenocarcinoma 77 (95.06%) 65 (87.84%) 0.198 31 (91.18%) 32 (100.00%) 0.239

Mucinous adenocarcinoma 2 (2.47%) 2 (2.70%) 0 (0.00%) 0 (0.00%)

squamous carcinoma 2 (2.47%) 7 (9.46%) 3 (8.82%) 0 (0.00%)

Metastatic lymph nodes There were 19 (23.46%) 3 (4.05%) <0.001 12 (35.29%) 1 (3.13%) 0.001

None 62 (76.54%) 71 (95.95%) 22 (64.71%) 31 (96.87%)

Signs	Category	Training Set (n = 155)	P	Validation Set (n = 66)	P
Age	Mean value (years)	60.42±10.13	61.38±10.18	0.558	62.09±10.18	62.50±11.83	0.880
Gender	Male	32 (39.51%)	36 (48.65%)	0.262	16 (47.06%)	14 (43.75%)	0.810
	Female	49 (60.49%)	38 (51.35%)		18 (52.94%)	18 (56.25%)
Size	Mean diameter (mm)	19.91±5.95	16.88±5.58	0.001	18.95±7.17	15.82±4.52	0.039
CT value	Mean value (HU)	28.10 (26.10)	–62.65 (268.98)	<0.001	32.05 (18.68)	–26.80 (234.83)	0.001
Density	Pure ground glass	0 (0.00%)	3 (4.05%)	<0.001	1 (2.94%)	0 (0.00%)	<0.001
	Hybrid ground glass	19 (23.46%)	40 (54.05%)		3 (8.82%)	16 (50.00%)
	Solidity	62 (76.54%)	31 (41.90%)		30 (88.24%)	16 (50.00%)
Department (right middle lobe) (counted as upper lung)	Upper right lung	35 (43.21%)	35 (47.29%)	0.888	15 (44.12%)	15 (46.88%)	0.498
	Lower right lung	13 (16.05%)	13 (17.57%)		9 (26.47%)	6 (18.75%)
	Upper left lung	24 (29.63%)	20 (27.03%)		9 (26.47%)	7 (21.87%)
	Lower left lung	9 (11.11%)	6 (8.11%)		1 (2.94%)	4 (12.50%)
Shortest distance between lesion and pleura	DLP (mm)	0.01 (0.00)	1.00 (3.00)	<0.001	0.00 (0.25)	2.00 (4.83)	0.004
Relationship with pleura	Type I	6 (7.41%)	18 (24.32%)	<0.001	1 (2.94%)	9 (28.12%)	0.002
	Type II	2 (2.47%)	12 (16.22%)		1 (2.94%)	4 (12.50%)
	Type III	8 (9.88%)	8 (10.82%)		2 (5.88%)	4 (12.50%)
	Type IV	14 (17.28%)	16 (21.62%)		8 (23.53%)	9 (28.12%)
	Type V	23 (28.39%)	10 (13.51%)		13 (38.24%)	3 (9.38%)
	Type VI	28 (34.57%)	10 (13.51%)		9 (26.47%)	3 (9.38%)
Pleural depression sign	There were	65 (80.25%)	36 (48.65%)	<0.001	30 (88.24%)	15 (46.87%)	<0.001
	None	16 (19.75%)	38 (51.35%)		4 (11.76%)	17 (53.13%)
Degree of differentiation	High	49 (60.50%)	58 (78.38%)	0.041	14 (41.18%)	22 (68.75%)
	Medium	18 (22.22%)	7 (9.46%)		10 (29.41%)	9 (28.12%)
	Low	14 (17.28%)	9 (12.16%)		10 (29.41%)	1 (3.13%)
Pathological typing	Adenocarcinoma	77 (95.06%)	65 (87.84%)	0.198	31 (91.18%)	32 (100.00%)	0.239
	Mucinous adenocarcinoma	2 (2.47%)	2 (2.70%)		0 (0.00%)	0 (0.00%)
	squamous carcinoma	2 (2.47%)	7 (9.46%)		3 (8.82%)	0 (0.00%)
Metastatic lymph nodes	There were	19 (23.46%)	3 (4.05%)	<0.001	12 (35.29%)	1 (3.13%)	0.001
	None	62 (76.54%)	71 (95.95%)		22 (64.71%)	31 (96.87%)

The measurement data conforming to normal distribution were expressed as mean±standard deviation and compared between groups by t-test; data conforming to skewed distribution were expressed as median (interquartile spacing) and compared between groups by Mann Whitney U-test; count data were expressed as frequency (percentage) and compared by chi-square test or Fisher exact test. p < 0.05 indicates that the differences were statistically significant.

3.2 Texture feature analysis and model construction

The mRMR and LASSO feature analysis of all 804 texture parameters in groups A and B identified 10 texture features that were of predictive significance (Fig. 2). These 10 features are:

F₁ –Wavelet_LLL_glcm_ClusterProminence,

F₂ –Wavelet_LLH_glcm_ClusterShade,

F₃ –Wavelet_LHH_glszm_GrayLevelVariance,

F₄ –Wavelet_HLL_glcm_Idmn,

F₅ –Wavelet_HLH_glrlm_LowGrayLevelRunEmphasis,

F₆ –Wavelet_LLH_firstorder_Variance,

F₇ –Wavelet_HLH_glszm_LargeAreaLowGrayLevelEmphasis,

F₈ –Wavelet_HLH_glszm_SmallAreaHighGrayLevelEmphasis,

F₉ –Wavelet_LHL_glcm_ClusterProminence,

F₁₀ –Wavelet_LHH_glszm_LowGrayLevelZoneEmphasis.

Then, a following texture feature-based prediction model was built to compute a likelihood score of the testing case being positive for VPI as: $\begin{matrix} Radscore = - 0.486 \times F_{1} + 0.351 \times F_{2} + 0.268 \times F_{3} + 0.243 \times F_{4} + 0.226 \times F_{5} \\ + 0.047 \times F_{6} + 0.307 \times F_{7} + 0.104 \times F_{8} + 0.085 \times F_{9} + 0.041 \\ \times F_{10} + 0.112 \end{matrix}$

The ROC curve was then applied to test the diagnostic efficacy and predictive performance of the model, and the results showed that the AUC of the texture feature model was 0.845 and 0.824 in the training and validation sets, respectively (p < 0.05) as shown in Fig. 2 and Table 2.

Table 2
Evaluation of diagnostic efficacy of texture feature model, CT imaging feature model, and joint prediction model

Models Training Set (n = 155) Validation Set (n = 66)

AUC (95% CI) Sensitivity Specificity AUC (95% CI) Sensitivity Specificity

Model_Texture 0.845 (0.778∼0.898) 72.97% 88.89% 0.824 (0.711∼0.907) 81.25% 76.47%

Model_CT 0.801 (0.729∼0.860) 59.46% 86.42% 0.882 (0.779∼0.948) 87.50% 79.41%

Model_Combined 0.900 (0.842∼0.943) 71.62% 92.59% 0.894 (0.794∼0.957) 71.87% 94.12%

Models	Training Set (n = 155)	Validation Set (n = 66)
Model_Texture	0.845 (0.778∼0.898)	72.97%	88.89%	0.824 (0.711∼0.907)	81.25%	76.47%
Model_CT	0.801 (0.729∼0.860)	59.46%	86.42%	0.882 (0.779∼0.948)	87.50%	79.41%
Model_Combined	0.900 (0.842∼0.943)	71.62%	92.59%	0.894 (0.794∼0.957)	71.87%	94.12%

3.3 Results of analysis of CT imaging signs and constructing models in two groups

There was good agreement between the two physicians for the assessment of CT sign evaluation, ROI setting and feature extraction (Consistency test ICC = 0.812 to 0.995).

Comparative analysis of CT imaging characteristics between the two groups showed that the differences between the two groups in mean diameter (19.91±5.95 mm vs. 16.88±5.58 mm), mean CT (28.10 vs. 62.65), density, shortest interpleural distance DLP (0.01 mm vs. 1.00 mm), typing of relationship with pleura, pleural depression signs (80.25% vs 48.65%), degree of differentiation and lymph node metastasis (23.46% vs 4.05%) were statistically significant (P < 0.05).

The above CT imaging features with statistically significant differences were included in the logistic regression analysis to construct the CT imaging feature model, and the results showed that the mean diameter, density, typing of relationship with pleura, and presence of lymph node metastasis were independent predictors of the CT imaging model. The ROC curve was used to test the diagnostic efficacy of the model, and the results showed that the AUC of the CT morphological characteristics model was 0.801 and 0.882 in the training set and validation set, respectively (P < 0.05) as shown in Table 2 and Fig. 3.

Fig. 3

Comparison of ROC curves of the prediction models.

3.4 Establishment of joint prediction model and evaluation of diagnostic efficacy

The indicators with diagnostic significance (10 texture features and 8 CT imaging features) from the results of the above texture features and CT imaging features analysis were further included in the logistic regression analysis to construct a joint prediction model, and the ROC curve was used to test the diagnostic efficacy of the model, and the results showed that the AUC of the joint model for the training and validation sets were 0.900, 0.894, respectively; see Table 2, Fig. 3.

In the training and validation sets, the results showed that the AUC of the joint prediction model was significantly higher than that of the texture feature model and the CT imaging model, respectively (0.900 vs 0.801 vs 0.845; 0.894 vs 0.824 vs 0.882); the specificity of the joint prediction model was also higher than that of the texture feature model and the CT imaging model (92.59% vs. 88.89% vs. 86.42%; 94.12% vs. 76.47% vs. 79.41%); see Table 2.

4 Discussion

VPI of lung cancer is an important predictor of postoperative recurrence, pleural implantation and pulmonary metastasis [3, 12] and is crucial for the prognosis of early-stage NSCLC [13], especially [2] for lung cancers ≤3.0 cm in diameter with negative lymph nodes. Zhao et al. [14] showed that the 5-year survival rate of NSCLC with VPI peripheral type was 10% –30% lower than that without VPI. Because of the internal heterogeneity of malignant tumors, there are limitations in previous relevance studies on the pure CT sign features to predict VPI in lung cancer.

In studies of the relationship between CT signs and VPI [6 , 16], most of them have not integrated the analysis of lung adenocarcinoma presenting various positions in relation to the pleura, and some of them have only analyzed lung cancer closely adjacent to the pleura, and for non-adjacent pleural lung cancer, the pleural depression sign is considered to be of great value in identifying VPI [17, 18]. In this study, by retrospective analysis of CT features and joint imaging quantitative parameters between VPI positive and negative patients, CT imaging, texture feature and the joint models were constructed using single-factor and multi-factor regression analysis. Our results showed that the AUC of CT imaging model, texture feature model and joint prediction model were (0.801, 0.845, 0.900) in training set and (0.882, 0.824, 0.894) in validation set, respectively. The AUC of the joint model was significantly higher than that of each independent model and had better predictive efficacy. Meanwhile, the morphological model showed that mean diameter, density, RAP, and presence of lymph node metastasis were all independent predictors of VPI.

It has shown that CT features had predictive efficacy for VPI and its prognosis. Deng et al. [19] studied 403 cases of NSCLC ≤3.0 cm in diameter and found that the mean size of the VPI-negative patients was significantly smaller than that of the VPI-positive patients (P < 0.001), which was consistent with the findings of Tanju et al. [20]. Zhang et al. [21] studied a large sample and confirmed that VPI-negative stage T1 NSCLC patients had better prognosis than VPI-positive patients, with 3-year survival rates of 76.7% –80.6% and 71.7%, respectively. Thus, VPI may directly affect the prognosis and survival of patients.

The density and solid component of the tumor are significantly correlated with VPI. Qi et al. [15] showed higher incidence of VPI in solid lung adenocarcinoma than in lung adenocarcinoma with ground glass density (54% vs 30%). Yip et al. [22] found that the incidence of VPI was significantly higher in solid nodules than in sub-solid nodules. Other related studies showed that the likelihood of VPI was greater in sub-solid nodules with >50% solid component; in contrast, for ≤3.0 cm purely ground glass nodules pGGN, VPI almost did not occur due to weak invasion and difficulty in breaking through the elastic fiber layer [15 , 23], suggesting that tumor density is also one of the independent predictors of VPI.

The shortest tumor pleural distance DLP is an important risk factor for VPI in peripheral lung cancer [21]. Deng et al. [19] found that the ratio of DLP ≤1.0 cm in the VPI-positive patients (69.8%) was significantly higher than that in the VPI-negative patients (48.5%, P < 0.001). It was hypothesized that the closer the tumor was to the pleura, the more likely VPI would occur. Due to the different pleural anatomical spaces, previous studies [16, 19] classified RAP into proximal and non-proximal pleura; the data in this study did not refine the classification of RAP relationship, and the DLP data were obtained by multiplanar reconstruction MPR technique to be true and objective, when the regression model was constructed for analysis in this study, they were excluded because DLP could not be used as an independent predictive model, but the results showed that DLP was significantly different in both the training set and validation set, while the DLP in the VPI-positive patients was significantly higher than that in the VPI-negative patients (0.94 mm vs. 2.43 mm), which is consistent with previous literature results, suggesting that the shorter the DLP, the higher the incidence of VPI.

For nodule-pleura relationship RAP, many studies have shown that no PI is important in determining VPI. Hsu et al. [16] showed that PI with soft tissue component at the pleural end is highly suggestive of VPI, and its formation is thought to be the result of structural destruction of lung lobules, thickening of interlobular septum and lymphatic vessel obstruction with edema, tumor cell infiltration along lymphatic vessels or inflammatory cell infiltration and fibrosis. Kim et al. [24] found that PI is a predictor of VPI and postoperative independent risk factor for poor prognosis, which is more prone to local recurrence, pleural and distant metastases, and a decrease in 5-year survival rate of approximately 22.4% was found in PI patients compared to those without PI. In this study, types IV–VI were classified as PI, and the results showed significant differences between the PI positive and negative patients (P < 0.001); the rates of PI of types IV, V, and VI were (19.13%, 31.30%, 32.17%, 23.58%, 12.26%, and 12.26%) between the PI positive and negative patients, respectively, with little difference between groups for PI of type IV, which was considered to be related to the selection of sample size and weighting; this study is consistent with the above findings.

Studies have shown a significant correlation between VPI and lymph node metastasis. Kudo et al. [25] showed that the probability of lymph node metastasis in peripheral type lung cancer with VPI ≤3.0 cm in diameter were higher than in the group without VPI lung cancer (9.2% vs 0.9%, P < 0.001). The visceral pleura distributes a large network of lymphatic vessels on the lung surface, which can enter the parabronchial lymph nodes and eventually the mediastinal lymph nodes through the intrapulmonary lymphatic system, and at the same time, as the VPI grade increases, it is prone to intra-thoracic dissemination and distant metastasis, which is an important factor leading to poor patient prognosis [2, 13]. However, imaging assessment of early-stage lung cancer has limitations due to the presence of lymph node micrometastasis [26] and the stealthiness of the intrapulmonary lymph nodes. Our study found significant difference in lymph node metastasis between the VPI positive and negative patients (26.96% vs. 3.77%), indicating that lymph node metastasis was an independent risk factor for predicting VPI, which was consistent with the above findings and suggests the need for postoperative adjuvant chemotherapy and close follow-up.

Radiomics reveals the intrinsic heterogeneous characteristics of tumors through high-throughput extraction of characteristic data from images, quantitative analysis of the distribution and association of pixels and voxel grayscale in image data, and deep excavation of their fine structure and change patterns. Yang et al. [7] studied the imaging features of lung nodules combined with clinical risk factor models, and their AUCs for benign and malignant differential diagnosis reached 0.935 (training group), 0.817 (validation group), respectively. Yagi et al. [27] studied the invasive imaging features of early lung cancer and showed higher values of kurtosis, skewness and homogeneity characteristic parameters of preinvasive carcinoma and micro-infiltrating adenocarcinoma compared with invasive adenocarcinoma, and multivariate analysis showed that CT 90th percentile and entropy were independent predictors with high accuracy and AUC of 0.90.

The current imaging characteristic parameters for peripheral type T1 stage lung cancer of VPI is not much studied. The study of Yuan et al. [28] on predicting VPI and prognosis by radiomics of 327 cases of ≤3.0 cm lung adenocarcinoma found that the accuracy, sensitivity, and specificity of imaging features predicting VPI reached 90.5%, 90.6%, and 93.2%, respectively; and the percentile 10%, wavEnll S 2 S 0 1 SumAverage of imaging features were independent risk factors for predicting VPI.

In this study, the texture features of lesions were extracted on CT plain lung windows, and a total of 804 feature parameters were extracted, resulting in the screening of 10 wavelet texture features containing grey level co-occurrence matrix (glcm), gray-level run-length matrix (glrlm), gray-level size zone matrix (glszm) and firstorder_Variance, covering first-order to high-order texture features, which can better reflect the spatial heterogeneity of lesions. Diagnostic models were developed using multivariate logistic regression, and the results showed that the AUCs of the texture feature models for diagnosing VPI in the training and validation sets were 0.845 and 0.824, respectively, and the sensitivity and specificity of both were 72.97% and 88.89%, 81.25% and 76.47%, respectively. Among them, Cluster Prominence has the best diagnostic efficacy with an AUC value of 0.741 and sensitivity and specificity of 63.48% and 80.37, respectively. Cluster Prominence belongs to one of the features of glcm, which indicates a measure of glcm skewness and inhomogeneity, and the higher its value, the stronger the asymmetry of the mean value. The large number of texture features extracted in this study indicates that the internal structure and spatial heterogeneity of tumors were closely related to VPI, but the diagnostic efficacy of individual features in predicting the occurrence of VPI was not particularly high, which might be related to the fact that individual features could only reflect part of the internal texture features of tumors, and multiple texture features need to be combined to improve the diagnostic accuracy.

The results of the ROC curve plotted in this study revealed that the AUC value of the CT feature model for predicting VPI was 0.801 in the training group, with a high diagnostic specificity of 86.42%, but a low sensitivity of 59.46%. The texture feature model had an AUC value of 0.845 for the diagnosis of VPI in training group, with a sensitivity of 72.97% and a specificity of 88.89%, which had better diagnostic efficacy. The AUC value of the two models combined was 0.900, and the corresponding sensitivity and specificity were 71.62% and 92.59%, respectively, and the diagnostic efficacy of which were significantly improved compared with the previous two models. Therefore, CT scan-based radiomics combined with CT features have great potential value in predicting ≤3.0 cm early non-small cell lung cancer with visceral pleural invasion; the link between VPI and texture features or deep learning will be further explored in the future to provide a quantitative basis for clinical decision making.

Despite encouraging results, we also recognize that this study has several limitations. First, the sample size was insufficient, especially for samples after subdivision into 6 types by pleural relationship RAP. Second, this study used two-dimensional image information for texture feature extraction, which has limitations relative to 3D texture feature extraction of the whole volume of lung cancer. Last, the statistical treatment of data related to postoperative survival and prognosis of patients with VPI was not done for this group data, and it is hoped that multicenter studies will be conducted in this aspect in the future.

In conclusion, for ≤3.0 cm peripheral type early-stage NSCLC, careful analysis of CT image features, combined with the information of 2D image texture features, can effectively predict the presence of VPI and guide the clinical formulation of reasonable treatment plans when suspicious VPI exists on preoperative imaging.

Footnotes

Acknowledgments

Not applicable.

References

Zeng

, Chen

, Zheng

et al., Changing cancer survival in China during –15: A pooled analysis of 17 population-based cancer registries, Lancet Glob Health 6(5) (2018), e555–e556.

Tian

, Pei

, Zheng

et al., Effect of visceral pleural invasion on the prognosis of patients with lymph node negative non-small cell lung cancer, Thorac Cancer 8(2) (2017), 97–105.

Neri

, Menju

, Sowa

et al., Prognostic impact of microscopic vessel invasion and visceral pleural invasion and their correlations with epithelial-mesenchymal transition, cancer stemness, and treatment failure in lung adenocarcinoma, Lung Cancer 128 (2019), 13–19.

Rami-Porta

, Asamura

, Travis

W.D.

et al., Lung cancer - major changes in the American Joint Committee on Cancer eighth edition cancer staging manual, CA Cancer J Clin 67(2) (2017), 138–155.

Bradbury

, Sivajohanathan

, Chan

et al., Postoperative adjuvant systemic therapy in completely resected non-small-cell lung cancer: A systematic review, Clin Lung Cancer 18(3) (2017), 259–273.e8.

Imai

, Minamiya

, Ishiyama

et al., Use of CT to evaluate pleural invasion in non-small cell lung cancer: Measurement of the ratio of the interface between tumor and neighboring structures to maximum tumor diameter, Radiology 267(2) (2013), 619–626.

Yang

, He

, Wang

et al., CT-based radiomics signature for differentiating solitary granulomatous nodules from solid lung adenocarcinoma, Lung Cancer 125 (2018), 109–114.

Zhang

, Cai

, Wang

et al., CT and clinical characteristics that predict risk of EGFR mutation in non-small cell lung cancer: A systematic review and meta-analysis, Int J Clin Oncol 24(6) (2019), 649–659.

Sun

, Hu

, Ge

et al., Radiomics study for predicting the expression of PD-L1 in non-small cell lung cancer based on CT images and clinicopathologic features, J Xray Sci Technol 28(3) (2020), 449–459.

10.

Starkov

, Aguilera

T.A.

, Golden

D.I.

et al., The use of texture-based radiomics CT analysis to predict outcomes in early-stage non-small cell lung cancer treated with stereotactic ablative radiotherapy, Br J Radiol 92(1094) (2019), 20180228.

11.

Yuan

The application of using Radiomics for phenotype of lung adenocarcinoma, Nanjing Medical University (2018).

12.

Jiwangga

, Cho

, Kim

et al., Recurrence pattern of pathologic stage I lung adenocarcinoma with visceral pleural invasion, Ann Thorac Surg 103(4) (2017), 1126–1131.

13.

Liu

Q.X.

, Deng

X.F.

, Zhou

et al., Visceral pleural invasion impacts the prognosis of non-small cell lung cancer: A meta-analysis, Eur J Surg Oncol 42(11) (2016), 1707–1713.

14.

Zhao

L.L.

, Xie

H.K.

, Zhang

L.P.

et al., Visceral pleural invasion in lung adenocarcinoma ≤3cm with ground-glass opacity: A clinical, pathological and radiological study, J Thorac Dis 8(7) (2016), 1788–1797.

15.

L.P.

, Li

X.T.

, Yang

et al., Multivariate analysis of pleural invasion of peripheral non-small cell lung cancer-based computed tomography features, J Comput Assist Tomogr 40(5) (2016), 757–762.

16.

Hsu

J.S.

, Han

I.T.

, Tsai

T.H.

et al., Pleural tags on CT scans to predict visceral pleural invasion of non-small cell lung cancer that does not abut the pleura, Radiology 279(2) (2016), 590–596.

17.

Ahn

S.Y.

, Park

C.M.

, Jeon

Y.K.

et al., Predictive CT features of visceral pleural invasion by T1-sized peripheral pulmonary adenocarcinomas manifesting as subsolid nodules, Am J Roentgenol 209(3) (2017), 561–566.

18.

Ebara

, Takashima

, Jiang

et al., Pleural invasion by peripheral lung cancer: Prediction with three-dimensional CT, Acad Radiol 22(3) (2015), 310–319.

19.

Deng

H.Y.

, Li

, Luo

et al., Novel biologic factors correlated to visceral pleural invasion in early-stage non-small cell lung cancer less than 3cm, J Thorac Dis 10(4) (2018), 2357–2364.

20.

Tanju

, Erus

, Selçukbiricik

et al., Level of pleural invasion effects on prognosis in lung cancer, Tumori 105(2) (2019), 155–160.

21.

Zhang

, Zhang

J.T.

, Li

W.F.

et al., Visceral pleural invasion in T1 tumors (≤3cm), particularly T1a, in the eighth tumor-node-metastasis classification system for non-small cell lung cancer: A population-based study, J Thorac Dis 11(7) (2019), 2754–2762.

22.

Yip

, Ma

, Flores

R.M.

et al., Survival with parenchymal and pleural invasion of non-small cell lung cancers less than 30mm, J Thorac Oncol 14(5) (2019), 890–902.

23.

Zhao

, Wang

J.W.

, Yang

et al., CT diagnosis of pleural and stromal invasion in malignant subpleural pure ground-glass nodules: An exploratory study, Eur Radiol 29(1) (2019), 279–286.

24.

Kim

H.J.

, Cho

J.Y.

, Lee

Y.J.

et al., Clinical significance of pleural attachment and indentation of subsolid nodule lung cancer, Cancer Res Treat 51(4) (2019), 1540–1548.

25.

Kudo

, Saji

, Shimada

et al., Impact of visceral pleural invasion on the survival of patients with non-small cell lung cancer, Lung Cancer 78(2) (2012), 153–160.

26.

Travis

W.D.

, Brambilla

, Burke

A.P.

et al., Introduction to the world health organization classification of tumors of the lung, pleura, thymus, and heart, J Thorac Oncol 10(9) (2015), 1240–1242.

27.

Yagi

, Yamazaki

, Ohashi

et al., HRCT texture analysis for pure or part-solid ground-glass nodules: Distinguishability of adenocarcinoma in situ or minimally invasive adenocarcinoma from invasive adenocarcinoma, Jpn J Radiol 36(2) (2018), 113–121.

28.

Yuan

, Liu

J.Y.

, Zhang

et al., Prognostic impact of the findings on thin-section computed tomography in stage I lung adenocarcinoma with visceral pleural invasion, Sci Rep 8(1) (2018), 4743.