Sage Journals: Discover world-class research

Abstract

Background

Differential diagnosis of single-nodule pulmonary metastasis (SNPM) and second primary lung cancer (SPLC) in patients with colorectal cancer (CRC) prior to lung surgery is relatively complex. Radiomics is an emerging technique for image information analysis, while it has not yet been applied to construct a differential diagnostic model between SNPM and SPLC in patients with CRC. In the present study, we aimed to extract radiomics signatures from thin-section computed tomography (CT) images of the chest. These radiomics signatures were combined with clinical features to construct a composite differential diagnostic model.

Method

A total of 91 patients with CRC, including 66 patients with SNPM and 25 patients with SPLC, were enrolled in this study. Patients were randomly assigned to the training cohort (n = 63) and validation cohort (n = 28) at a ratio of 7 to 3. Moreover, 107 radiomics features were extracted from the chest thin-section CT images. The least absolute shrinkage and selection operator (LASSO) regression was used to filter these features, and clinical features were screened by univariate analysis. The screened radiomics and clinical features were combined to construct a multifactorial logistic regression composite model. The receiver operating characteristic (ROC) curves were adopted to evaluate the models, and the corresponding nomograms were created.

Results

A series of 6 radiomics characteristics was screened by LASSO. After univariate logistic regression analysis, the composite model finally included 4 radiomics features and 4 clinical features. In the training cohort, the area under the curve scores of ROC curves were 0.912 (95% confidence interval [CI]: 0.813-0.969), 0.884 (95% CI: 0.778-0.951), and 0.939 (95% CI: 0.848-0.984) for models derived from radiomics, clinical, and combined features, respectively. Similarly, these values were 0.756 (95% CI: 0.558-0.897), 0.888 (95% CI: 0.711-0.975), and 0.950 (95% CI: 0.795-0.997) in the validation cohort, respectively.

Conclusions

We constructed a model for differential diagnosis of SNPM and SPLC in patients with CRC using radiomics and clinical features. Moreover, our findings provided a new assessment tool for patients with CRC in the future.

Keywords

colorectal cancer single-nodule pulmonary metastasis second primary lung cancer CT images radiomics machine learning

Introduction

As one of the leading causes of tumor-related death, colorectal cancer (CRC) remains a significant public health challenge globally.^1,2 Because of the presence of hematogenous metastases, pulmonary metastases (PMs) are common lung malignant lesions in patients with CRC, accounting for approximately 10% to 20% of patients. Resection of PMs is currently thought to improve progression-free survival in patients who can tolerate it. In terms of number, single-nodule pulmonary metastasis (SNPM) can be better surgically resected. Patients with CRC with SNPM are not uncommon in clinical practice.^3–7 In contrast to SNPM, second primary lung cancer (SPLC) is a pulmonary malignancy found simultaneously or after the primary tumor.^8,9 SPLC in patients with CRC is not rare due to underlying immune deficiencies, gene mutations, carcinogen exposure, lifestyle, or cancer treatment.^10,11 Before pulmonary surgery, their differential diagnosis in clinical practice is challenging. To precisely distinguish them, pathomorphological observation and immunohistochemical detection are often achieved by bronchoscopy or percutaneous needle biopsy,¹² while these operations are not applicable in some patients. Therefore, an effective pre-operative examination is required to detect and identify potential metastases early.

In patients with CRC, it is necessary to evaluate lung lesions to confirm initial staging, treatment strategy, and follow-up plan. There is a consensus that imaging of the chest, including X-rays, computed tomography (CT), and positron emission tomography (PET)-CT, is a vital pre-operative and postoperative monitoring tool for patients with CRC.^13–16 To date, there have been studies using imaging and clinical features to establish predictive models in the early diagnosis of PMs.^17–19 However, it is still tricky for the readout results of CT and PET-CT images to be entirely consistent with the pathological findings.²⁰ In recent years, radiomics has gradually attracted people's attention. In radiomics, CT, PET-CT, and other images are converted into high-dimensional, mineable data through high-throughput extraction of multiple quantitative features, followed by subsequent data analysis to build predictive models for decision support.^21,22 Radiomics has been applied to the prediction of pre-operative staging, the monitoring of distant metastases, and the differentiation of pulmonary nodules in CRC.^23–26 Deep learning, combined with radiomics, has been shown to be a valuable tool for evaluating patients with CRC with pulmonary metastases (PMs).²⁷ There is also a growing body of research focused on using deep learning and related data augmentation algorithms to analyze lung malignancies.^28,29

In clinical diagnosis, differential diagnosis of SNPM and SPLC is still relatively complex. Therefore, we extracted the radiomics signatures of lung lesions in patients with CRC and combined the radiomics model with the clinical model to establish a reliable composite differential diagnostic model. The main contributions are as follows:

Radiomics features of CT images were applied to isolated nodules in the lungs of patients with CRC, and it provided a good differential diagnosis between SNPM and SPLC. With models built through multiple machine learning approaches, radiomics can effectively assist in solving this problem.

This study also established a clinical image model for comparison and combination of radiomics models. The radiomics model in this study had comparable diagnostic efficacy to the traditional clinical image model.

A combined model was constructed by integrating radiomics features with clinical features. The diagnostic efficacy of this model was better than that of existing methods.

Materials and Methods

Patient Selection

This research was a retrospective study. A total of 6401 patients with CRC who received oncotherapy at the First Affiliated Hospital of Soochow University from January 2013 to June 2021 were reviewed. The inclusion criteria of our study were: (1) CRC with histopathological confirmation under surgical resection or colonoscopy; (2) only pathological diagnosis of adenocarcinoma; (3) relevant CT images and medical history were available; and (4) patients with regular re-examination after tumor treatment. The exclusion criteria were as follows: (a) the number of pulmonary nodules > 1 when pulmonary lesions were first detected on chest thin-section CT images; (b) patients with primary cancers other than CRC and lung cancer; and (c) image sequences that could not be sketched due to image quality problems. Eventually, 2 categories of patients (Table 1) were identified: (A) patients with SNPM (n = 66) and (B) patients with SPLC (n = 25). All patient details were de-identified in the study process.

Table 1.

Characteristics of Patients With CRC with Different Pulmonary Nodules.

Characteristics	SNPM (n = 66)	SPLC (n = 25)	t/χ2	P
Gender			1.885	.170
Male	40	19
Female	26	6
Age (median ± SD, years)	64.15 ± 9.44	64.41 ± 9.75	−0.255	.849
CEA			0.017	.898
Positive	15	6
Negative	51	19
CA125			1.315	.252
Positive	3	15
Negative	22	51
ITP, months			4.490	.034*
≤ 12	40	21
＞ 12	26	4
Maximum diameter (median ± SD, mm)	25.16 ± 15.113	19.70 ± 13.101	1.701	.092
Spicule sign			6.292	.012*
Positive	23	16
Negative	43	9
Vacuole sign			7.629	.006*
Positive	9	10
Negative	57	15
GGO			22.044	.000*
Positive	1	9
Negative	65	16

* Statistically significant, P < .05.

Abbreviations: CA125, carbohydrate antigen 125; CEA, carcino-embryonic antigen; GGO, ground-glass opacity; SNPM, single-nodule pulmonary metastases; SPLC, second primary lung cancer; ITP, interval to pulmonary nodules.

Acquisition of Image Data and Clinical Data

All patients underwent chest thin-section CT examinations in the supine position using the GE LightSpeed VCT scanner or GE BrightSpeed Elite 16 scanner (GE Healthcare, Milwaukee, USA). Scans were performed from the lung apex to the lung base using the following parameters: tube voltage 120 kV, tube current automatic mA, helical pitch 0.938, rotation speed 0.6 s, slice thickness, and spacing 5 mm. CT images were reconstructed with a slice thickness of 1.00 to 1.25 mm. The images were reviewed by 2 chest radiologists with more than 10 years of experience together. If there was a disagreement between the 2 radiologists, a third radiologist with 20 years of experience conducted a further review. During this procedure, the following image features were recorded, including maximum lesion diameter, spicule sign, and vacuolation sign. In addition, some clinical data of these patients were collected through the clinical information system, including age, gender, interval to pulmonary nodules (ITP), carcino-embryonic antigen (CEA), and carbohydrate antigen 125 (CA125). The above data were summarized in a table consisting of clinical characteristics.

CT Image Analysis

Chest thin-section CT images were downloaded from the hospital's picture archiving and communication system to an offline workstation for texture analysis. The operator (Y.Y., a graduate student) implemented lesion segmentation using 3D Slicer software (v4.13.0; National Institutes of Health, NIH, USA). The visible boundaries of the lesions were manually outlined based on the regions of interest in each axial tomography. The lesion boundaries were segmented into 2-dimensional regions and then converted into volumes of interest (VOIs) for 3-dimensional (3D) data. The areas of blood vessels, air, normal pulmonary tissue, and surrounding organs were excluded as much as possible in this process. An expert radiologist (DSM, with 20 years of experience in pectoral and abdominal CT) then finally validated the process and revised it. After finishing delineation, the 3D data of VOIs and image sequences (NII format) were transferred to FeAture Explorer software (FAE, V 0.3.6) on Python (3.7.6) (https://github.com/salan668/FAE) for feature extraction and analysis.³⁰ After preprocessing of data, radiomics features with missing values and noncontiguous variables were excluded. A total of 107 features were extracted finally. The number of radiomics features based on feature classes were as follows and shown in Table 2: (i) 18 for first-order feature, (ii) 14 for shape feature, (iii) 24 for gray level co-occurrence matrix feature, (iv) 16 for gray level size zone matrix (GLSZM) feature, (v) 16 for gray level run length matrix feature, (vi) 5 for neighborhood gray-tone difference matrix feature, and (vii) 14 for gray level dependence matrix (GLDM) feature.

Table 2.

Characteristics of the Training and Validation Cohorts.

Method	Texture feature parameters
First-order features (n = 18)	10Percentile, 90Percentile, Energy, Entropy, Interquartile Range, Kurtosis, Maximum, Mean, Mean Absolute Deviation, Median, Minimum Range, Robust Mean Absolute Deviation, Root Mean Squared, Skewness, Total Energy, Uniformity, Variance
Shape features (n = 14)	Elongation, Flatness, Least Axis Length, Major Axis Length, Maximum2DDiameterColumn, Maximum2DDiameterRow, Maximum2DDiameterSlice, Maximum3DDiameter, Mesh Volume, Minor Axis Length, Sphericity, Surface Area, Surface Volume Ratio, Voxel Volume
Gray level co-occurrence matrix features (n = 24)	Autocorrelation, Cluster Prominence, Cluster Shade, Cluster Tendency, Contrast, Correlation, Difference Average, Difference Entropy, Difference Variance, Id, Idm, Idmn, Idn, Imc1, Imc2, Inverse Variance, Joint Average, Joint Energy, Joint Entropy, MCC, Maximum Probability, Sum Average, Sum Entropy, Sum Squares
Gray level size zone matrix features (n = 16)	Gray Level NonUniformity, Gray Level NonUniformity Normalized, Gray Level Variance, High Gray Level Zone Emphasis, Large Area Emphasis, Large Area High Gray Level Emphasis, Large Area Low Gray Level Emphasis, Low Gray Level Zone Emphasis, Size Zone NonUniformity, Size Zone NonUniformity Normalized, Small Area Emphasis, Small Area High Gray Level Emphasis, Small Area Low Gray Level Emphasis, Zone Entropy, Zone Percentage, Zone Variance
Gray level run length matrix feature (n = 16)	Gray Level NonUniformity, Gray Level NonUniformity Normalized, Gray Level Variance, High Gray Level Run Emphasis, Long Run Emphasis, Long Run High Gray Level Emphasis, Long Run Low Gray Level Emphasis, Low Gray Level Run Emphasis, Run Entropy, Run Length NonUniformity, Run Length NonUniformity Normalized, Run Percentage, Run Variance, Short Run Emphasis, Short Run High Gray Level Emphasis, Short Run Low Gray Level Emphasis
Neighborhood gray-tone difference matrix feature (n = 5)	Busyness, Coarseness, Complexity, Contrast, Strength
Gray level dependence matrix feature (n = 14)	Dependence Entropy, Dependence NonUniformity, Dependence NonUniformity Normalized, Dependence Variance, Gray Level NonUniformity, Gray Level Variance, High Gray Level Emphasis, Large Dependence Emphasis, Large Dependence High Gray Level Emphasis, Large Dependence Low Gray Level Emphasis, Low Gray Level Emphasis, Small Dependence Emphasis, Small Dependence High Gray Level Emphasis, Small Dependence Low Gray Level Emphasis

Radiomics Feature Selection and Radiomics Model Development

After extraction of features, computer-generated random datasets were used to assign 70% of datasets to the training set and 30% of the datasets to the independent validation set. Consequently, 63 cases were selected as the training dataset (46/17 = SNPM/SLPC), while the other 28 cases constituted the independent validation dataset (20/8 = SNPM/SLPC).

To remove the unbalance of the training dataset, up-samples were conducted by repeating random cases to make positive/negative samples balance. Then, the normalization was applied to the feature matrix (Z score, mean). Since the dimension of feature space was high, the linear models were penalized with the L1 norm or the least absolute shrinkage and selection operator (LASSO), Pearson correlation coefficient, and principal component analysis to select features before building the model. After this process, the dimension of the feature space was reduced, and each feature was independent of the others. Lastly, the analysis of variance, relief, and recursive feature elimination were utilized for feature selections. Multiple methods were used as the classifier, including the support vector machine, linear discriminant analysis, auto-encoder, random forests, logistic regression (LR), LR via LASSO, ada-boost, decision tree, Gaussian process, and naive Bayes. Besides, a 10-fold cross-validation was applied to the training dataset to determine the hyper-parameter (such as the number of features) of the model (Supplemental Figure 1 and Table 1). The hyper-parameters were set according to the model performance on the validation dataset. The above operations were implemented by FAE software and Python 3.0 (https://www.python.org).

The performance of the radiomics models was evaluated by the receiver operating characteristic (ROC) curve analysis, Youden index, and F1 scores. Subsequently, the accuracy, sensitivity, specificity, positive predictive value, and negative predictive value of each model were calculated.

Statistics Analysis in Clinical and Combined Models

The clinical features were age, gender, ITP, CEA, CA125, maximum lesion diameter, spicule sign, vacuolation sign, and ground-glass opacity (GGO). Statistical analyses were performed using SPSS software version 26.0 (SPSS Inc., Chicago, IL, USA) and Medcalc software version 20 (https://www.medcalc.org/). The nominal variables were analyzed by the chi-square test or Fisher's exact test. The continuous variables with abnormal distribution were analyzed by the Mann-Whitney test, while the continuous variables with normal distribution were analyzed by t-test. Moreover, univariate logistic analysis was employed to investigate the discriminative radiomics and clinical features. Multivariate analyses were performed using the logistic regression analysis model in the training and validation cohorts. A P < .05 was considered statistically significant (Figure 1). Based on ground truth data, clinical models were developed in the training and validation sets. The model was manifested by ROC curves, and the differential diagnostic performance of the model was assessed by the area under the curve (AUC). The reporting of this study conformed with STROBE guidelines.³¹

Figure 1.

The flow diagram of this study. Abbreviation: LASSO, least absolute shrinkage and selection operator.

Results

Patient Clinical Characteristics

Table 1 summarizes the clinical characteristics of the patients in the training and validation cohorts. A total of 91 patients were included in this study. Among the enrolled patients, 66 (72.5%) were patients with SNPM, and 25 (27.5%) were patients with SPLC. Among these patients, there were no statistically significant differences in clinical features between the training and validation cohorts (Table 2). ITP (P = .041), spicule sign (P = .014), vacuolation sign (P = .008), and GGO (P = .001) were significant correlates of PM in the univariate logistics regression analysis (Table 3).

Table 3.

Characteristics of Patients With CRC in the Training Cohort and Validation Cohort.

Characteristics	Total (n = 91)	Training (n = 63)	Validation (n = 28)	t/χ²	P
Gender				0.771	.380
Male	32	24	8
Female	59	39	20
Label				0.025	.876
SNPM	66	46	20
SPLC	25	17	8
Age, median ± SD, years	62.42 ± 11.832	61.79 ± 11.519	63.82 ± 12.611	−0.753	.454
CEA				0.062	.804
Positive	21	15	7
Negative	70	48	21
CA125				0.694	.405
Positive	18	11	7
Negative	73	52	21
ITP, months				0.012	.991
≤ 12	61	42	19
＞ 12	30	21	9
Maximum diameter, median ± SD, mm	21.20 ± 13.816	22.90 ± 15.533	17.36 ± 7.752	1.789	.077
Spicule sign				0.843	.359
Positive	39	25	14
Negative	52	38	14
Vacuole sign
Positive	19	13	6	0.007	.931
Negative	72	50	22
GGO				0.003	.955
Positive	10	7	3
Negative	81	56	25

Extraction and Selection of Radiomics Features

A total of 107 radiomics features were extracted from the CT images (Table 2). Multiple radiomics models were built using different normalization, dimensionality reduction screening, and machine learning methods. Several simple models with good validation efficacy were selected based on the validation results. These models are listed in Table 4. In addition, a cutoff value of 6 features was chosen based on the model's performance, and it was used to compare the performance of radiomics models in different machine learning pipelines (Figure 2).

Figure 2.

AUCs of different datasets using 10 machine learning algorithms by different methods. Abbreviations: AUC, area under the curve; ANOVA, analysis of variance; KW, Kruskal-Wallis; PCA, principal component analysis; PCC, Pearson correlation coefficient; RFE, recursive feature elimination.

Table 4.

Multiple Machine Learning Pipelines for Radiomics Models.

	AUC	95% CI	Std	Acc	Youden Index	Sen	Spe	PPV	NPV
Mean_PCA_ANOVA_6_LR	0.9130	0.8408-0.9853	0.0368	0.9048	0.7954	0.913	0.8824	0.9545	0.7895
Mean_PCA_ANOVA_6_LRLasso	0.8964	0.8174-0.9754	0.0403	0.8889	0.7737	0.8913	0.8824	0.9535	0.7500
Mean_PCA_ANOVA_5_GP	0.8785	0.7862-0.9708	0.0471	0.8254	0.6867	0.8043	0.8824	0.9487	0.6250
Mean_PCC_RFE_9_LDA	0.8747	0.7695-0.9798	0.0537	0.8730	0.7148	0.8913	0.8235	0.9318	0.7368
Zscore_PCC_RFE_9_LDA	0.8747	0.7695-0.9798	0.0537	0.8730	0.7148	0.8913	0.8235	0.9318	0.7368
Mean_PCA_ANOVA_7_AE	0.8696	0.7825-0.9566	0.0444	0.8095	0.665	0.7826	0.8824	0.9474	0.6000
Mean_PCA_RFE_4_LRLasso	0.8696	0.7816-0.9575	0.0449	0.7778	0.6215	0.7391	0.8824	0.9444	0.5556
Mean_PCA_RFE_4_LR	0.8670	0.7772-0.9568	0.0458	0.8254	0.6125	0.8478	0.7647	0.907	0.6500
Mean_PCA_ANOVA_5_SVM	0.8606	0.7683-0.9529	0.0471	0.8254	0.6867	0.8043	0.8824	0.9487	0.6250
Mean_PCA_KW_4_LR	0.8606	0.7696-0.9516	0.0464	0.8095	0.6650	0.7826	0.8824	0.9474	0.6000
Mean_PCC_Relief_8_LDA	0.8593	0.7616-0.9570	0.0498	0.7460	0.6151	0.6739	0.9412	0.9688	0.5161
Zscore_PCC_Relief_8_LDA	0.8593	0.7616-0.9570	0.0498	0.7460	0.6151	0.6739	0.9412	0.9688	0.5161
Zscore_PCA_RFE_4_LRLasso	0.8555	0.7617-0.9493	0.0478	0.7619	0.5997	0.7174	0.8824	0.9429	0.5357
Mean_PCA_KW_4_LRLasso	0.8542	0.7614-0.9470	0.0473	0.8254	0.6496	0.8261	0.8235	0.9268	0.6364
Mean_PCC_RFE_3_AE	0.8529	0.7532-0.9526	0.0509	0.7778	0.6215	0.7391	0.8824	0.9444	0.5556
Mean_PCA_RFE_3_GP	0.8504	0.7382-0.9626	0.0572	0.8254	0.6867	0.8043	0.8824	0.9487	0.6250
Mean_PCA_KW_4_SVM	0.8453	0.7483-0.9422	0.0495	0.7619	0.6368	0.6957	0.9412	0.9697	0.5333

Abbreviations: 95% CI, 95% confidence interval; Acc, accuracy; AE, auto-encoder; ANOVA, analysis of variance; AUC, area under the curve; GP, Gaussian process; KW, Kruskal-Wallis; LASSO, least absolute shrinkage and selection operator; LDA, linear discriminant analysis; LR, logistic regression; LRLASSO, logistic regression via LASSO; NPV, negative predictive value; PCA, principal component analysis; PPV, positive predictive value; RFE, recursive feature elimination; Sen, sensitivity; Spe, specificity; Std, standard error; SVM, support vector machine.

Diagnostic Performance of Individual Models

As a frequently used method, the LASSO was chosen to filter features using dimensionality reduction in building the combined model when keeping other processes unchanged. Six radiomics features were screened, and again these were included in the univariate logistics regression analysis, including Glszm_LargeAreaHighGrayLevelEmphasis (P = .336), Gldm_DependenceEntropy (P = .002), Firstorder_InterquartileRange (P = .019), Firstorder_90Percentile (P = .001), Firstorder_Maximum (P = .098), and Shape_Sphericity (P = .000). We obtained 4 significantly associated radiomics features (P < .05) (Table 5), included them in a multifactorial logistic regression analysis along with clinical characteristics, and modeled them accordingly.

Table 5.

Univariate Analysis for Models by Logistic Regression.

Variables	B	Std	Wald	P	Exp(B)	95.0% CI for Exp (B)
Variables	B	Std	Wald	P	Exp(B)	Lower	Upper
Glszm_LargeAreaHighGrayLevelEmphasis	0.000	0.000	0.924	.336	1.000	1.000	1.000
Gldm_DependenceEntropy	−4.218	1.361	9.599	.002*	0.015	0.001	0.212
Firstorder_InterquartileRange	−0.006	0.003	5.547	.019*	0.994	0.988	0.999
Firstorder_90Percentile	0.007	0.002	11.564	.001*	1.007	1.003	1.010
Firstorder_Maximum	0.001	0.001	2.736	.098	1.001	1.000	1.003
Shape_Sphericity	9.755	2.483	15.440	.000*	17240.360	132.843	2237454.995
Gender	−0.722	0.532	1.843	.175	0.486	0.171	1.378
Age	−0.006	0.020	0.096	.756	0.994	0.955	1.034
CEA	−0.071	0.553	0.017	.898	0.931	0.315	2.752
CA125	0.769	0.682	1.270	.260	2.157	0.567	8.209
ITP	1.227	0.601	4.173	.041*	3.412	1.051	11.080
Maximum diameter	−0.027	0.016	2.676	.102	0.974	0.943	1.005
Spicule sign	−1.201	0.490	6.002	.014*	0.301	0.115	0.786
Vacuole sign	−1.440	0.543	7.025	.008*	0.237	0.082	0.687
GGO	−3.599	1.090	10.894	.001*	0.027	0.003	0.232

* Statistically significant, P < .05.

Abbreviations: 95% CI, 95% confidence interval; CA125, carbohydrate antigen 125; CEA, carcino-embryonic antigen; GGO ground-glass opacity; ITP, interval to pulmonary nodules; Std, standard error.

Figure 3A and B show the AUCs of ROC curves for 3 models (radiomics, clinical, and combined models) individually in the training and validation cohorts, respectively. In the training cohort, the AUC scores for differentially diagnosing SNPM and SPLC were 0.912 (95% confidence interval [CI]: 0.813-0.969), 0.884 (95% CI: 0.778-0.951), and 0.939 (95% CI: 0.848-0.984) for models derived from radiomics features, clinical features, and combined features, respectively. Similarly, in the validation cohort, the AUC scores of 3 models were 0.756 (95% CI: 0.558-0.897), 0.888 (95% CI: 0.711-0.975), and 0.950 (95% CI: 0.795-0.997), respectively (Table 6).

Figure 3.

Performance of different models. (A) Training cohort and (B) validation cohort. Abbreviation: AUC, area under the curve.

Table 6.

ROC Analysis for Models by Logistic Regression.

Groupings	AUC	Std	95% CI
Training cohort
Radiomics model	0.912	0.0362	0.813-0.969
Clinical model	0.884	0.0436	0.778-0.951
Combined model	0.939	0.0276	0.848-0.984
Validation cohort
Radiomics model	0.756	0.1240	0.558-0.897
Clinical model	0.888	0.0654	0.711-0.975
Combined model	0.950	0.0388	0.795-0.997

Abbreviations: 95% CI, 95% confidence interval; AUC, the area under the curve; Std, standard error.

In addition, we constructed nomograms based on the results of multivariate logistic regression analysis for the training (Figure 4A) and validation cohorts (Figure 4B). For each patient with CRC with an SNPM in the lung, a corresponding score was obtained based on radiomics and clinical variables. The probability of a diagnosis of SNPM or SPLC could be assessed by calculating the total score.

Figure 4.

Nomogram for differential diagnosis of SNPM and SPLC. (A) Training cohort and (B) validation cohort. Abbreviations: SNPM, single-nodule pulmonary metastasis; SPLC, second primary lung cancer.

Discussion

The type of lung lesion for patients with CRC presenting with pulmonary nodules of unknown provenance determines the subsequent surgery and chemotherapy treatment plan.³² Except for benign lung nodules that require no further treatment, metastatic lung lesions are often resected through minimally invasive surgery.³³ In contrast, primary lung cancer involves complete lung lobe resection, and further mediastinal lymph node dissection is required.^34,35 Despite the vital value for postoperative lung surgery,^36,37 the differential diagnosis of individual pulmonary malignant nodules has been poorly studied due to the small number of patients compared with the overall number. Radiomics and deep learning models are increasingly being used to evaluate suspicious lung nodules and assist in the decision-making process. For instance, Ren et al have proposed a hybrid framework that enhances various classification models to address the task of lung cancer classification, which has achieved better performance compared to other state-of-the-art methods.^28,29 To the best of our knowledge, no studies have combined radiomics features with clinical features to establish a differential diagnostic model between SNPM and SPLC (Figure 5).

Figure 5.

Pulmonary malignant lesions in patients with CRC, all located in the right upper lung but presenting differently. The light green overlay represents the actual segmentation area in the figure. (A) Single pulmonary nodule with an irregular shape, multiple spicules, and vacuole signs in a 62-year-old male patient. Postoperative pathology was found to be SNPM. (B) Pulmonary mass lesions with a relatively regular shape in a 73-year-old male patient. Postoperative pathology was SPLC. (C) Nearly circular pulmonary nodule in a 75-year-old male patient, a common manifestation of PMs. Postoperative pathology confirmed SNPM. (D) The single lung nodule was observed as GGO in this image. Postoperative pathological findings in this 63-year-old male patient's lung CRC proved to be SPLC. Abbreviations: CRC, colorectal cancer; GGO, ground-glass opacity; PMs, pulmonary metastases; SNPM, single-nodule pulmonary metastasis; SPLC, second primary lung cancer.

In the present study, we extracted radiomics features from chest thin-section CT images of different natures of malignant lung lesions in patients with CRC and established a radiomics model. The radiomics features in this model included Gldm_DependenceEntropy, Firstorder_InterquartileRange, Firstorder_90Percentile, and Shape_Sphericity. The model had moderate diagnostic efficacy for the differential diagnosis, whether the lung lesions were lung metastases or SPLC. This result was also confirmed by the method of multiple radiomics study cohorts. Compared with our clinical models using the spicule sign, vacuole sign, GGO, and ITP, the 2 had very similar diagnostic efficacy. Furthermore, a composite model combining the 2 was superior to either the clinical model or the radiomics model alone.

With our newly developed clinical model, we believed that the spicule sign and vacuole sign could all be used as diagnostic references to distinguish SNPM from SPLC. This conclusion was consistent with the findings of several previous studies.^38–41 Pre-existing research has also concluded that spicule sign, vacuole sign, and GGO are significant factors in the univariate analysis, while spicule sign and vacuole sign are not significant factors in the differential diagnosis of SNPM and SPLC by multivariate analysis.⁴² This conclusion is not consistent with some of our findings. The reason might be attributed to the difference in the subjective opinion of the person reading the CT images, further illustrating the necessary introduction of other types of features, such as radiomics, histopathology, immunohistochemistry, and genomics.

In previous studies, ITP is considered an important factor affecting the prognosis of PMs in patients with CRC because it reflects the biology of the tumor. An ITP > 12 months generally represents a worse prognosis (4). In the differential diagnosis of SNPM and SPLC, some scholars have obtained by subgroup analysis that the division of ITP into > 3 and ≤ 3 can help distinguish the 2. However, the time cutoff value taken is still controversial in the known studies.⁴³ In our present study, ITP was classified traditionally as > 12 (n = 30) and ≤ 12 (n = 61). A univariate analysis was performed to obtain P = .041 (P < .05). Although we considered ITP as a significant independent predictor to identify the 2 and included it in the clinical and combined models, the choice of time cutoff values in ITP needs to be further analyzed in future studies.

Serum CEA is a common tumor marker for monitoring CRC metastasis and recurrence. It may also be relevant for predicting the occurrence of CRC lung metastasis.⁴⁴ However, some previous studies have shown that serum CEA is not a significant independent factor in the differential diagnosis of pulmonary lesions. Even in patients with PMs, most patients show negative serum CEA.^18,45 Similar results were obtained in our study. This finding indicated that tumor markers, such as serum CEA, might not be effective in the differential diagnosis of PM and SPLC in patients with CRC.

Reproducibility and robustness are vital in radiomics analysis.⁴⁶ In the present study, all chest CT images were realized in the same center using almost similar acquisition and reconstruction protocols. In addition to the 10-fold cross-validation, we used different downscaling, screening, normalization, and modeling approaches in our radiomics analysis to support the soundness of this process.

There were several limitations in our study. First, this work was carried out based on a single-center retrospective study, which might lead to potential selection bias. External validation of the model should be conducted with multicenter patient data, when possible, to further evaluate its effectiveness and generalizability. Second, although we had the largest sample size of any known study in this direction, the sample size was still relatively small for machine learning. Prospective trials should be conducted to recruit a sufficient number of patients with CRC with SNPM or SPLC, in order to further evaluate the clinical usefulness of the radiomics differential diagnosis model. Third, the type of clinical information we collected was limited, which might make us unable to uncover potentially significant independent predictors. To improve the generalizability of the combined model, future studies should consider adding information from multiple sources, such as metabolic data from PET/CT, genomics data, and pathomics data.

Conclusions

A radiomics model was developed using machine learning to analyze lung CT images, which demonstrated good performance in differentiating SNPM and SPLC in patients with CRC. The radiomics model exhibited comparable efficacy to the clinical image model that included the spicule sign, vacuole sign, GGO, and ITP. By combining the radiomics model with clinical image features, a composite model was generated, which could enhance diagnostic accuracy and assist in making appropriate clinical decisions for patients with CRC.

Supplemental Material

sj-png-1-tct-10.1177_15330338231175735 - Supplemental material for The Differential Diagnostic Value of Radiomics Signatures Between Single-Nodule Pulmonary Metastases and Second Primary Lung Cancer in Patients with Colorectal Cancer

Supplemental material, sj-png-1-tct-10.1177_15330338231175735 for The Differential Diagnostic Value of Radiomics Signatures Between Single-Nodule Pulmonary Metastases and Second Primary Lung Cancer in Patients with Colorectal Cancer by Yu Yu, Jiaqing Tan, Yi Yang, Bin Zhang, Xiaodong Yao, Shibiao Sang and Shengming Deng in Technology in Cancer Research & Treatment

Supplemental Material

sj-docx-2-tct-10.1177_15330338231175735 - Supplemental material for The Differential Diagnostic Value of Radiomics Signatures Between Single-Nodule Pulmonary Metastases and Second Primary Lung Cancer in Patients with Colorectal Cancer

Supplemental material, sj-docx-2-tct-10.1177_15330338231175735 for The Differential Diagnostic Value of Radiomics Signatures Between Single-Nodule Pulmonary Metastases and Second Primary Lung Cancer in Patients with Colorectal Cancer by Yu Yu, Jiaqing Tan, Yi Yang, Bin Zhang, Xiaodong Yao, Shibiao Sang and Shengming Deng in Technology in Cancer Research & Treatment

Supplemental Material

sj-pdf-3-tct-10.1177_15330338231175735 - Supplemental material for The Differential Diagnostic Value of Radiomics Signatures Between Single-Nodule Pulmonary Metastases and Second Primary Lung Cancer in Patients with Colorectal Cancer

Supplemental material, sj-pdf-3-tct-10.1177_15330338231175735 for The Differential Diagnostic Value of Radiomics Signatures Between Single-Nodule Pulmonary Metastases and Second Primary Lung Cancer in Patients with Colorectal Cancer by Yu Yu, Jiaqing Tan, Yi Yang, Bin Zhang, Xiaodong Yao, Shibiao Sang and Shengming Deng in Technology in Cancer Research & Treatment

Footnotes

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Ethics Statement

This study is a retrospective study. Only the clinical data of the patients were collected, and the treatment plan of the patients was not intervened. The researchers have tried their best to protect the information provided by patients from disclosing their personal privacy. This study was reviewed and approved by the Ethics Committee of the First Affiliated Hospital of Soochow University (Approval number: 2022.302).

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This research was funded by the National Natural Science Foundation of China (No. 81601522), Medical Youth Talent Project of Jiangsu Province (No. QNRC2016749), Gusu Health Talent Program (No. GSWS2020013), Suzhou People's Livelihood Science and Technology Project (No. SYS2019038), Project of State Key Laboratory of Radiation Medicine and Protection, Soochow University (No. GZK1202127), and the open Foundation of Nuclear Medicine Laboratory of Mianyang Central Hospital (No. 2021HYX023 and 2021HYX029).

ORCID iD

Yu Yu

Supplemental Material

Supplemental material for this article is available online.

Abbreviations

References

GBD 2017 Colorectal Cancer Coll. Aborators. The global, regional, and national burden of colorectal cancer and its attributable risk factors in 195 countries and territories, 1990–2017: a systematic analysis for the Global Burden of Disease Study 2017. Lancet Gastroenterol Hepatol. 2019;4(12):913‐933. Doi: 10.1016/S2468-1253(19)30345-0

Chen

Ren

, et al. Participation and yield of a population-based colorectal cancer screening programme in China. Gut. 2019;68(8):1450‐1457. Doi: 10.1136/gutjnl-2018-317124

Sakamaki

Ishida

Tanaka

. Prognosis of patients with recurrence after pulmonary metastasectomy for colorectal cancer. Gen Thorac Cardiovasc Surg. 2020;68(10):1172‐1178. Doi: 10.1007/s11748-020-01368-5

Cho

Song

Yang

Jheon

. Prognostic factors of pulmonary metastasis from colorectal carcinoma. Interact Cardiovasc Thorac Surg. 2013;17(2):303‐307. Doi: 10.1093/icvts/ivt164

Moorcraft

Ladas

Bowcock

Chau

. Management of resectable colorectal lung metastases. Clin Exp Metastasis. 2016;33(3):285‐296. Doi: 10.1007/s10585-015-9774-6

Cho

Kim

Namgung

, et al. The prognostic importance of the number of metastases in pulmonary metastasectomy of colorectal cancer. World J Surg Oncol. 2015;13(1):222. Doi: 10.1186/s12957-015-0621-7

Tan

Lopes Gde

Jr Sim

. How uncommon are isolated lung metastases in colorectal cancer? A review from database of 754 patients over 4 years. J Gastrointest Surg. 2009;13(4):642‐648. Doi: 10.1007/s11605-008-0757-7

Feller

Matthes

Bordoni

, et al. The relative risk of second primary cancers in Switzerland: a population-based retrospective cohort study. BMC Cancer. 2020;20(1):51. Doi: 10.1186/s12885-019-6452-0

Quint

Park

Iannettoni

. Solitary pulmonary nodules in patients with extrapulmonary neoplasms. Radiology. 2000;217(1):257‐261. Doi: 10.1148/radiology.217.1.r00oc20257

10.

Tanaka

Ogawa

Mizusawa

, et al.. Second primary cancers and recurrence in patients after resection of colorectal cancer: an integrated analysis of trials by Japan Clinical Oncology Group: JCOG1702A. Jpn J Clin Oncol. 2021;51(2):185‐191. Doi: 10.1093/jjco/hyaa184

11.

Mery

Pappas

Bueno

, et al. Relationship between a history of antecedent cancer and the probability of malignancy for a solitary pulmonary nodule. Chest. 2004;125(6):2175‐2181. Doi: 10.1378/chest.125.6.2175

12.

Yuan

Yang

, et al. Expert consensus on multidisciplinary therapy of colorectal cancer with lung metastases (2019 edition). J Hematol Oncol. 2019;12(1):16. Doi: 10.1186/s13045-019-0702-0

13.

Benson

Venook

Al-Hawary

, et al. Colon cancer, version 2.2021, NCCN clinical practice guidelines in oncology. J Natl Compr Canc Netw. 2021;19(3):329‐359.

14.

Benson

Venook

Al-Hawary

, et al. NCCN guidelines insights: rectal cancer, version 6.2020. J Natl Compr Canc Netw. 2020;18(7):806‐815. Doi: 10.6004/jnccn.2020.0032

15.

Pomerri

Pucciarelli

Maretto

, et al. Significance of pulmonary nodules in patients with colorectal cancer. Eur Radiol. 2012;22(8):1680‐1686. Doi: 10.1007/s00330-012-2431-4

16.

Bamba

Itabashi

Kameoka

. Value of PET/CT imaging for diagnosing pulmonary metastasis of colorectal cancer. Hepatogastroenterology. 2011;58(112):1972‐1974. Doi: 10.5754/hge11256

17.

Ohtaki

Shimizu

Nagashima

, et al. Clinical and radiological discrimination of solitary pulmonary lesions in colorectal cancer patients. World J Surg. 2018;42(4):1161‐1170. Doi: 10.1007/s00268-017-4243-9

18.

Jung

Kim

Ryu

Paik

Hwang

. Indeterminate pulmonary nodules in colorectal cancer. World J Gastroenterol. 2015;21(10):2967‐2972. Doi: 10.3748/wjg.v21.i10.2967

19.

Huang

Chen

Chang

Tsai

Huang

. Surgical outcomes of second primary lung cancer after the extrapulmonary malignancy. J Cancer Res Clin Oncol. 2020;146(12):3323‐3332. Doi: 10.1007/s00432-020-03310-x

20.

Guerrera

Renaud

Schaeffer

, et al. Low accuracy of computed tomography and positron emission tomography to detect lung and lymph node metastases of colorectal cancer. Ann Thorac Surg. 2017;104(4):1194‐1199. Doi: 10.1016/j.athoracsur.2017.05.002

21.

Gillies

Kinahan

Hricak

. Radiomics: images are more than pictures, they are data. Radiology. 2016;278(2):563‐577. Doi: 10.1148/radiol.2015151169

22.

Mayerhoefer

Materka

Langs

, et al. Introduction to radiomics. J Nucl Med. 2020;61(4):488‐495. Doi: 10.2967/jnumed.118.222893

23.

Liang

Huang

, et al. The development and validation of a CT-based radiomics signature for the preoperative discrimination of stage I-II and stage III-IV colorectal cancer. Oncotarget. 2016;7(21):31401‐31412. Doi: 10.18632/oncotarget.8919

24.

Wang

Yin

Gao

. Radiomics and radiogenomics in evaluation of colorectal cancer liver metastasis. Front Oncol. 2022;11(1):689509. Doi: 10.3389/fonc.2021.689509

25.

Wang

Huang

, et al. A clinical-radiomics nomogram for the preoperative prediction of lung metastasis in colorectal cancer patients with indeterminate pulmonary nodules. Eur Radiol. 2019;29(1):439‐449. Doi: 10.1007/s00330-018-5539-3

26.

Jin

Wang

Zhang

, et al.

Solitary pulmonary lesion in patients with history of malignancy: primary lung cancer or metastatic cancer?

Ann Surg Oncol. 2018;25(5):1237‐1244. Doi: 10.1245/s10434-018-6360-6

27.

Wang

Dai

Gong

, et al. Development of a novel combined nomogram model integrating deep learning-pathomics, radiomics and immunoscore to predict postoperative outcome of colorectal cancer lung metastasis patients. J Hematol Oncol. 2022;15(1):11. Doi: 10.1186/s13045-022-01225-3

28.

Ren

Zhang

Wang

. A hybrid framework for lung cancer classification. Electronics (Basel). 2022;11(10):1614. Doi: 10.3390/electronics1010000

29.

Ren

Zhang

Wang

. LCDAE: data augmented ensemble framework for lung cancer classification. Technol Cancer Res Treat. 2022;21(1):15330338221124372. Doi: 10.1177/15330338221124372

30.

Song

Zhang

, et al. Feature explorer (FAE): a tool for developing and comparing radiomics models. PLoS One. 2020;15(8):e0237587. Doi: 10.1371/journal.pone.0237587

31.

von Elm

Altman

Egger

, et al. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies. Ann Intern Med. 2007;147(8):573‐577. Doi: 10.7326/0003-4819-147-8-200710160-00010

32.

Rena

Davoli

Boldorini

, et al. The solitary pulmonary nodule in patients with previous cancer history: results of surgical treatment. Eur J Surg Oncol. 2013;39(11):1248‐1253. Doi: 10.1016/j.ejso.2013.08.014

33.

Laisaar

Vooder

Umbleja

. Thoracoscopic resection of a solitary pulmonary nodule in patients with a known history of malignancy. Thorac Cardiovasc Surg. 2008;56(7):418‐421. Doi: 10.1055/s-2008-1038347

34.

Chan

Gaur

, et al. Management of the solitary pulmonary nodule. Arch Pathol Lab Med. 2017;141(7):927‐931. Doi: 10.5858/arpa.2016-0307-RA

35.

Zellweger

Abdelnour-Berchtold

Krueger

Ris

Perentes

Gonzalez

. Surgical treatment of pulmonary metastasis in colorectal cancer patients: current practice and results. Crit Rev Oncol Hematol. 2018;127(1):105‐116. Doi: 10.1016/j.critrevonc.2018.05.001

36.

Cho

Kim

Namgung

, et al. The prognostic importance of the number of metastases in pulmonary metastasectomy of colorectal cancer. World J Surg Oncol. 2015;13(1):222. Doi: 10.1186/s12957-015-0621-7

37.

Tanaka

Ogawa

Mizusawa

38.

Lee

Jeong

Kim

. Differentiation of primary lung cancer from solitary lung metastasis in patients with colorectal cancer: a retrospective cohort study. World J Surg Oncol. 2021;19(1):28. DOI: 10.1186/s12957-021-02131-7

39.

Gou

Chen

, et al.

Clinical characteristics of patients with solitary pulmonary mass after radical treatment for primary cancers: pulmonary metastasis or second primary lung cancer?

Cancer Invest. 2013;31(6):397‐403. Doi: 10.3109/07357907.2013.800092

40.

Zhong

Liu

Wang

, et al. A predictive model to differentiate between second primary lung cancers and pulmonary metastasis. Acad Radiol. 2022;29(Suppl 2):S137‐S144. DOI: 10.1016/j.acra.2021.05.015. Epub 2021 Jun 24.

41.

Nakadate

Sato

, et al. Predictors of primary lung cancer in a solitary pulmonary lesion after a previous malignancy. Gen Thorac Cardiovasc Surg. 2017;65(12):698‐704. Doi: 10.1007/s11748-017-0825-6

42.

Ohtaki

Shimizu

Nagashima

, et al. Clinical and radiological discrimination of solitary pulmonary lesions in colorectal cancer patients. World J Surg. 2018;42(4):1161‐1170. Doi: 10.1007/s00268-017-4243-9

43.

Chang

Tsai

, et al. The clincopathological variables to differentiate the nature of isolated pulmonary nodules in patients who received curative surgery for colorectal cancer. Asian J Surg. 2019;42(2):425‐432. Doi: 10.1016/j.asjsur.2018.08.002

44.

Nagata

Ishihara

Oba

, et al. Development and validation of a prediction model for organ-specific recurrences after curative resection of colon cancer. Dis Colon Rectum. 2019;62(9):1043‐1054. Doi: 10.1097/DCR.0000000000001430

45.

Pan

Zhao

Xiao

. Pulmonary metastasis in rectal cancer: a retrospective study of clinicopathological characteristics of 404 patients in Chinese cohort. BMJ Open. 2018;8(2):e019614. Doi: 10.1136/bmjopen-2017-019614

46.

Vallières

Zwanenburg

Badic

Cheze Le Rest

Visvikis

Hatt

. Responsible radiomics research for faster clinical translation. J Nucl Med. 2018;59(2):189‐193. Doi: 10.2967/jnumed.117.200501

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.02 MB

0.13 MB

0.21 MB