A machine learning-based prediction of the micropapillary/solid growth pattern in invasive lung adenocarcinoma with radiomics.

BACKGROUND: Micropapillary/solid (MP/S) growth patterns of lung adenocarcinoma are vital for making clinical decisions regarding surgical intervention. This study aimed to predict the presence of a MP/S component in lung adenocarcinoma using radiomics analysis.
METHODS: Between January 2011 and December 2013, patients undergoing curative invasive lung adenocarcinoma resection were included. Using the "PyRadiomics" package, we extracted 90 radiomics features from the preoperative computed tomography (CT) images. Subsequently, four prediction models were built by utilizing conventional machine learning approaches fitting into radiomics analysis: a generalized linear model (GLM), Naïve Bayes, support vector machine (SVM), and random forest classifiers. The models' accuracy was assessed using a receiver operating curve (ROC) analysis, and the models' stability was validated both internally and externally.
RESULTS: A total of 268 patients were included as a primary cohort, and 36.6% (98/268) of them had lung adenocarcinoma with an MP/S component. Patients with an MP/S component had a higher rate of lymph node metastasis (18.4% versus 5.3%) and worse recurrence-free and overall survival. Five radiomics features were selected for model building, and in the internal validation, the four models achieved comparable performance of MP/S prediction in terms of area under the curve (AUC): GLM, 0.74 [95% confidence interval (CI): 0.65-0.83]; Naïve Bayes, 0.75 (95% CI: 0.65-0.85); SVM, 0.73 (95% CI: 0.61-0.83); and random forest, 0.72 (95% CI: 0.63-0.81). External validation was performed using a test cohort with 193 patients, and the AUC values were 0.70, 0.72, 0.73, and 0.69 for Naïve Bayes, SVM, random forest, and GLM, respectively.
CONCLUSIONS: Radiomics-based machine learning approach is a very strong tool for preoperatively predicting the presence of MP/S growth patterns in lung adenocarcinoma, and can help customize treatment and surveillance strategies. 2021 Translational Lung Cancer Research. All rights reserved.

Introduction

Lung adenocarcinoma is established to be the most common subtype of lung cancer, with high heterogeneity in its molecular, pathological, and clinical aspects. Based on the histological pattern, a classification by the International Association for the Study of Lung Cancer/American Thoracic Society/European Respiratory Society further categorized invasive lung adenocarcinoma into the five following subtypes: lepidic, acinar, solid, papillary, and micropapillary (1). It is now required that each histological pattern and its percentage be recorded in the pathological reports. Many studies (2-4) have supported the histological classification of lung adenocarcinoma being correlated with patient’s prognosis, lymph node metastasis, and epidermal growth factor receptor (EGFR) mutation status. Patients with the micropapillary (MP)-predominant and solid (S)-predominant subtype are considered to have worse survival prognosis and higher recurrence rates. Studies also have shown that even a MP component of <5% has a significant prognostic impact on patient’s survival (5,6). Moreover, patient with an MP component >5% undergoing limited resection, but not lobectomy, would carry a higher risk of tumor recurrence compared with those having an MP component of <5% (7). Therefore, preoperative predictions of the presence of an MP/S component can provide important information for deciding a patient’s surgical strategies.

Radiomics (often called texture/shape analysis) can be used to characterize the medical images via quantitative image analysis; thus, it may represent a valuable source of information for lesion diagnosis and prognosis prediction (8). Machine learning, an artificial intelligence subfield, can utilize radiomics features to automatically infer decisions from the datasets with the purpose of developing a decision-making model. In regrading lung cancer, computed tomography (CT)-based radiomics are a practical, low-cost tool for noninvasively characterizing lung lesions and discerning between malignant and benign lung nodules (9), thus stratifying the risk of mediastinal lymph node metastasis (10) and discriminating invasive from indolent adenocarcinomas (11). However, highly accurate and reliable machine learning methods are still lacked in driving the success of radiomics applications in clinical care, and should be chosen appropriately (12).

Thus, in our study, we aimed to evaluate and compare the predictive value of radiomics-based machine learning models in the presence of an MP/S growth pattern of lung adenocarcinoma.

We present the following article in accordance with the TRIPOD reporting checklist (available at http://dx.doi.org/10.21037/tlcr-21-44).

Methods

This was a multicenter study with patients retrospectively included from three different institutions, including Shanghai Pulmonary Hospital, Ningbo No. 2 Hospital, and the Affiliated Hospital of Zunyi Medical College. The waiver of informed consent was approved by the institutional review board (IRB) of each participating hospital. The study was performed in accordance with the Declaration of Helsinki (as revised in 2013). The complete design of our study is illustrated in .

Figure 1

The complete design of this study. Tumor slide of micropapillary/solid (MP/S) pattern was stained with hematoxylin and eosin. Original magnification, ×20.

Study population

Patients undergoing resection for invasive lung adenocarcinoma from January 2011 to December 2013 at Shanghai Pulmonary Hospital, and from January 2013 to December 2014 in the other two hospitals were retrospectively recruited in this study. The inclusion criteria were as follows: (I) pathologically confirmed invasive lung adenocarcinoma; (II) complete preoperative thin-section CT image (1 mm) data with follow-up survival data; and (III) tumor size measured in CT images <3 cm. We excluded those patients with multifocal lesions, neoadjuvant chemotherapy or radiotherapy. The dataset from Shanghai Pulmonary Hospital with the most included patients (N=268) was used as the primary cohort for model development, and the datasets (N=193) from the other two hospitals were used as independent external validation cohorts.

All patients were followed up every 6 months for the first 3 years after surgery and every year thereafter. Follow-up examinations included thoracic CT scans, carcinoembryonic antigen (CEA) levels, abdominal CT scans or ultrasonography, bone scintigraphy, and cranial CT scans or magnetic resonance imaging (MRI). All patients were followed up for more than 5 years. Each included patient was restaged based on the 8th edition of lung cancer staging classification (13).

Histological evaluation

All formalin-fixed, paraffin-embedded (FFPE) tissue specimens were recut and stained with hematoxylin and eosin (HE). Two pathologists concurrently re-evaluated the slides using a multiheaded microscope, and discrepancies were resolved through discussion. Histological subtypes of lung adenocarcinoma were classified on the basis of the International Association for the Study of Lung Cancer/American Thoracic Society/European Respiratory Society multidisciplinary classification of lung adenocarcinoma (1). Each histological subtype component was recorded in 5% increments. Based on the presence or absence of an MP/S growth pattern (Figure S1A), the study population was split into two groups: (I) patients with MP/S components and (II) patients without.

Image acquisition and tumor segmentation

Thin-section CT scans of the lungs were obtained at full inspiration by the SOMATOM Definition AS scanner (64×0.625 mm detector, 1.0 pitch; Siemens, Germany) or a Brilliance 40 scanner (40×0.625 mm detector configuration, 0.4 pitch, Philips, The Netherlands) with 120 KVp of tube energy and 200 mAS of effective dose. A medium sharp reconstruction algorithm was utilized for image reconstruction with a section thickness of 1 mm and a 0.7-mm increment. All CT scans were performed without contrast medium. The identified CT scans were downloaded from the Picture Archiving and Communication Systems (PACS) (Figure S1B).

We used the open-source platform (3D-slicer, v4.9.0, www.slicer.org) to achieve tumor segmentation via the “segment editor” model. Tumors were delineated on the CT images in horizontal, sagittal, and coronal planes using a semi-automatic segmentation based on “level tracing” and “smoothing”. A radiologist and a thoracic surgeon then reviewed all tumor segmentations in consensus, and any discrepancies were resolved by additional correction.

Radiomics feature extraction and selection

“PyRadiomics”, an open-source package for standardizing the extraction of radiomics data (https://github.com/Radiomics/pyradiomics), was used to extract 90 radiomics features from the nonfiltered segmented tumor regions (ROIs). The extracted features can be classified into five categories: (I) shape (n=14); (II) intensity (n=19); (III) gray level co-occurrence matrix (GLCM; n=25); (IV) gray level run length matrix (GLRLM; n=16); and (V) gray level size zone matrix (GLSZM; n=16). Shape features quantified the three-dimensional (3D) geometry of the tumor, such as sphericity and surface area. Intensity features, also called first-order statistical features, are derived from the histogram of the tumor voxel, and can include the 90th percentile and standard deviation. GLCM, GLRLM, and GLSZM features characterize tumor texture and reflect the intratumor heterogeneity via an analysis of the spatial patterns of the tumor voxel. A detailed description and computing algorithms of the radiomics features are given in the Appendices (Appendix 1). Each radiomics feature was compared between the two subgroups, and the corresponding P value was calculated to test the significance. Multicollinearity of the radiomics features was visualized through a correlation heatmap. To reduce the collinearity of the extracted features and remove redundant information to avoid the overfitting issue, the most significant feature was selected from each category by identifying the feature with the lowest P value within each category before modelling (14).

Model building and validation

Based on the selected radiomics features, four popular machine learning methods were utilized to build the predictive models. These methods included the generalized linear model (GLM), Naïve Bayes, random forest, and support vector machine (SVM). In order to evaluate the model’s stability, a bootstrap resampling technique was used, and 100 random samples were iteratively selected with replacement from the original dataset. In each iteration, the model was constructed on the selected samples and validated using the remaining patients in the original dataset. The discrimination ability of model was assessed with the area under the curve (AUC) value. In addition, the clinical usefulness of these models was measured and compared by a decision curve analysis.

Statistical analysis

Statistical analyses were performed in R platform (R version 3.4.2). R package “ComplexHeatmap” was used for unsupervised cluster analysis. Package “e1071” was used for the implementation of Naïve Bayes and SVM classifiers, while the package “randomForest” was applied to develop the random forest model, both using the default parameter tuning with caret interface. A GLM via penalized maximum likelihood was fitted using the “glm” function with default parameters. Continuous and categorical variables were analyzed through the student’s t and chi-squared tests, respectively. Variables with a P value <0.1 in univariate analysis were entered into the multivariate logistic regression in a forward stepwise fashion. A Kaplan-Meier analysis was used to evaluate the survival outcomes of different groups, and differences among the survival curves were evaluated by a log-rank test. A two-sided P value <0.05 was considered statistically significant.

Results

Baseline characteristics and survival analysis

From January 2011 to December 2013, a primary cohort consisting of 268 patients with invasive lung adenocarcinoma at Shanghai Pulmonary Hospital was included in this study. The median age was 61 years (range, 30–87 years). Most of the patients were female (56%) and had no smoking history (79.1%). Furthermore, 10.1% (27/268) of patients had lymph node metastasis, among whom 20 were N1 positive, 17 were N2 positive, and 10 had both N1 and N2 metastasis. With regards to histological subtypes, 77 adenocarcinomas (28.7%) were revealed as having MP components, among which only 9 were MP-predominant. In terms of solid growth pattern, 33 adenocarcinomas contained solid components, and 25 of them were solid-predominant. Overall, 36.6% (98/268) of lung adenocarcinoma patients had an MP/S component with 12 presenting both components. The validation cohort consisted of 193 patients, and, with the exception of slightly larger lung tumors (mean: 2.47 versus 2.17 cm; P=0.003), there were no significant differences in baseline characteristics compared to that of patients in the primary cohort ().

Table 1

Baseline characteristics of patients with and without the micropapillary/solid pattern

Variables	Primary cohort (n=268)	Validation cohort (n=193)	P
Age, years	61 [30–87]	60 [35–82]	0.19
Sex, n (%)			0.43
Female	150 (56.0)	100 (51.8)
Male	118 (44.0)	93 (48.2)
Smoking history, n (%)			0.74
Smoker	56 (20.9)	37 (19.2)
Non-smoker	212 (79.1)	156 (80.8)
Lobular location, n (%)			0.54
RUL	84 (31.3)	57 (29.5)
RML	18 (6.7)	18 (9.3)
RLL	55 (20.5)	30 (15.5)
LUL	71 (26.5)	55 (28.5)
LLL	40 (14.9)	33 (17.1)
Predominant subtype, n (%)			0.21
Lepidic	90 (33.6)	73 (37.8)
Acinar	94 (35.1)	71 (36.8)
Papillary	50 (18.7)	28 (14.5)
Solid	25 (9.3)	14 (7.3)
Micropapillary	9 (3.4)	4 (2.1)
Lymph node metastasis	27 (10.1)	16 (8.1)	0.63
Size (cm)	2.47±0.88	2.17±0.84	0.003
With MP/S component, n (%)			1
No	170 (63.4)	122 (63.2)
Yes	98 (36.6)	71 (36.8)

RUL, right upper lobe; RML, right middle lobe; RLL, right lower lobe; LUL, left upper lobe; LLL, left lower lobe; CEA, carcinoembryonic antigen; MP, micropapillary; S, solid.

Clinical significance of the MP/S component

In univariate analysis, male patients with elevated CEAs and larger tumor sizes were more likely to be MP/S positive (Table S1). Among these three clinical factors, multivariate analysis demonstrated that sex was the only significant predictor of having an MP/S component. The presence of an MP/S component significantly correlated with tumor aggressive characteristics, including lymph node metastasis (18.4% versus 5.3%; P=0.001) and visceral pleural invasion (68.4% versus 55.3%; P=0.011). Overall and recurrence-free survival (OS and RFS, respectively) were compared between patients with MP/S components and those without. Log-rank tests revealed that the presence of an MP/S component was significantly associated with decreased OS and RFS in both the training (, both P<0.001) and the validation (Figure S2, P<0.001 and P=0.016) cohort. Even for the patients without lymph node metastasis, the prognostic effect of an MP/S growth pattern still remained significant (Figure S3).

Figure 2

Kaplan-Meier curves of survival comparison between patients with the MP/S growth pattern and those without. Recurrence-free survival (A) and overall survival (B) curves. MP/S, micropapillary/solid growth pattern.

Radiomics analysis and feature selection

Unsupervised cluster analysis demonstrated that the radiomics feature was significantly associated with histological subtypes. As shown in , three clusters of patients with similar radiomics pattern were significantly different in terms of predominant histological, MP-containing, and solid-containing subtypes, and the percentages of lung adenocarcinomas with an MP/S component were 62.5%, 1.9%, and 39.0%, respectively (P<0.001).

Figure 3

Radiomics heatmap with 268 patients in the y-axis and 90 radiomics features in the x-axis. Unsupervised cluster analysis revealed three clusters of patients with similar radiomics pattern, and the histological features of the lung adenocarcinomas were significantly different among these clusters in terms of pre-dominant, micropapillary-containing, and solid-containing subtypes.

The correlation heatmap (Figure S4) revealed that multicollinearity was present among these radiomics features. By identifying the features with the lowest p values within each radiomics category, the five radiomics features contributing the most to the diagnosis of MP/S-containing lung adenocarcinomas were selected to compensate for this multicollinearity problem. As shown in , by calculating the negative log10 of the p value, the five radiomics features (surface volume ratio, root mean squared, run entropy, joint average, and zone entropy) were top-ranked in the y-axis. The differences in these five radiomics features between patients with MP/S components and those without are summarized in . In an unsupervised cluster analysis of the five selected radiomics features, the percentage of patients with MP/S components was significantly different among the formed clusters in both the primary and external cohorts (Figure S5).

Figure 4

Model building and validation, and their predictive performance comparison. (A) Radiomics feature selection was accomplished by identifying the feature with the lowest P value within each category. (B) Evaluation of the model performance of the five modelling strategies by bootstrap resampling. The Y-axis indicates the AUC value in each iteration of sampling. (C) Model assessment with ROC curves in external validation.

Table 2

Selected radiomics features for predicting the micropapillary/solid pattern in lung adenocarcinoma

Feature class	Feature name	Student’s t test
		Without MP/S (n=170)	With MP/S (n=98)	P
Shape	Surface volume ratio	0.41±0.14	0.32±0.09	3.87E-08
First order	Root mean squared	694±174	827±96	2.12E-14
GLRLM	Run entropy	4.42±0.28	4.63±0.18	1.45E-13
GLCM	Joint average	9.65±2.75	11.91±1.69	5.06E-15
GLSZM	Zone entropy	5.87±0.47	6.17±0.23	1.92E-11

GLRLM, gray level run length matrix; GLCM, gray level co-occurrence matrix; GLSZM, gray level size zone matrix.

For the predictive model built by x, the probability of having an MP/S component can be simply expressed by the following equation: 1/[1+ exp (−0.0057× root mean squared − 1.93× zone entropy +25.4)]. The four machine leaning algorithms demonstrated similar model accuracy, with the Naïve Bayes model demonstrating the highest AUC value of 0.75 (95% CI: 0.65–0.85). The AUC values of each model in each iteration were all >0.6, which indicates good reliability of the selected radiomics features (). In the external validation, the AUC values were 0.70 (95% CI: 0.62–0.79), 0.72 (95% CI: 0.63–0.81), 0.73 (95% CI: 0.64–0.82), and 0.69 (95% CI: 0.61–0.77) for Naïve Bayes, SVM, random forest, and GLM, respectively (). The decision curve analysis indicated that there might be comparable net benefits among four proposed models (Figure S6).

Discussion

Given that the MP and S subtypes present as strong predictors for tumor recurrence and worse survival of lung adenocarcinoma, we can confidently assume that predicting the presence of an MP/S component could be vital for improving the prognosis of lung adenocarcinoma (5,15). In this study, we confirmed that patients with an MP/S component had a higher rate of lymph node metastasis and worse prognoses than those without. In addition, we demonstrated that four machine learning models based on five optimal radiomics features could achieve similar and relatively accurate MP/S prediction performance (highest AUC value of 0.75; 95% CI: 0.65–0.85), which confirms the good reliability of radiomics features.

Overwhelming evidence shows that the MP and solid subtypes present as strong predictors for worse prognosis, and the prognostic effect remains even with proportions <5% (5,15). Our results are consistent with these studies, which demonstrate that the presence of an MP/S growth pattern significantly correlated with lymph node metastasis (P=0.001) and visceral pleural invasion (P=0.011), as well as decreased OS and RFS. Furthermore, several reports (16-18) have also demonstrated that the MP pattern is positively associated with the EGFR mutation and anaplastic lymphoma kinase (ALK) fusion. It was also found that more than 85% of MP-predominant adenocarcinomas harbor EGFR driver mutations, and the MP growth pattern appears to be most likely correlated with a heterogeneous EGFR mutation, which acts as a mechanism of acquired resistance to EGFR tyrosine kinas inhibitors (TKIs) (18,19). Solid subtypes have otherwise been correlated with a high rate of KRAS mutation (20). Therefore, it is highly valuable to predict the presence of the MP/S growth pattern in helping to create personalized treatment and surveillance strategies for patients with lung adenocarcinoma.

Radiomics features can potentially decode the medical images to completely and non-invasively capture tumor phenotypic characteristics quantitatively, and has demonstrated promising predictive results for tumor invasiveness assessment (11), recurrence-free or overall survival (21,22), and treatment response (23). Regarding the histological growth pattern prediction, Park et al. (24) demonstrated that radiomics could differentiate three categories of the predominant subtypes in adenocarcinoma with favorable performance. However, the strong impact of MP/S component on poor prognosis needs to more accurately identify its presence, even those with a small portion. Song et al. (6) reported that a multivariable logistic model could identify two radiomics features, including the lower value for the minimum of whole pixel values and the lower value for the variance of positive pixel values, as imaging predictors for tumors with an MP component >5%. The unsupervised cluster analysis of our study also confirmed that the radiomics features are significantly associated with histological subtypes. However, the problem of multicollinearity leading to predictive model overfitting cannot be ignored. Jiang et al. (25) found that the random forest model built with reliable radiomics features (intraclass correlation coefficients ≥0.75) could obtain a good predictive performance in spreading through the air space in lung adenocarcinoma. Traverso et al. (26) proposed that machine learning–based radiomics methods may be useful for robust predictive model development. Thus in our study, after the pairwise comparison was performed to identify the most significant and predictive radiomics features from each category with the lowest P value, four popular machine learning methods were utilized to build the predictive models and achieve similar and stable performance in both internal and external validation. Our results indicated that these machine-learning methods turned out reasonably stable against data perturbation and hence they could be preferred for radiomics-based modeling. In addition, the five selected features were demonstrated to be highly reliable and could be utilized even when based on the CT images from different institutions, at different time points, and with different algorithms used for model construction.

This study also had some limitations which should be noted. First, both the primary and external validation cohorts were predominantly composed of female non-smoking patients with a relatively small sample size. Due to the imbalanced data seen in this study, a larger multicenter study would be needed for further validate the reproducibility and generalizability of our findings. Second, during radiomics feature extraction, the effects of CT scanner variability and inconsistent acquisition parameters were not considered. Further studies on radiomics should clarify the influence of the CT scanning setting on radiomics features, such as the application of the contrast medium. Lastly, the performance of the existed machine learning model for predicting the MP/S growth pattern was moderate and might be not yet suitable for clinical practice. The application of other novel ML-based predictive models or radiomics combined with deep learning in MP/S prediction may be worthwhile in future research.

In conclusion, radiomics features as applied to machine learning classification significantly correlate with histological subtypes of lung adenocarcinoma and seems to present a non-invasive, economical approach for MP/S growth pattern prediction.

The article’s supplementary files as