Background
Currently, lung cancer is the leading cause of cancer-related deaths, accounting for 23% of all cancer deaths [
1], and 80–85% of them are non-small cell lung cancer (NSCLC). With the broad application of thin-layer CT scanning technology and the continuous development of lung cancer screening programs, the detection rate of early lung adenocarcinoma with ground-glass nodules (GGNs) continues to increase [
2]. In many aspects, primary lung adenocarcinoma is considered as a very heterogeneous tumor with different histopathology and disease processes [
3]. According to the 2011 classification of adenocarcinoma proposed by the International Association for the Study of Lung Cancer, the American Thoracic Society, and the European Respiratory Society (IASLC/ATS/ERS) [
4], the most common patterns should be identified as the predominant growth patterns of invasive adenocarcinoma (IAC), including five subtypes: lepidic, acinar, papillary, micropapillary, and solid. The use of predominant growth patterns not only helps to classify IAC into subtypes but also serves as a prognostic indicator independent of clinical stage [
5,
6]. Among the first three most common growth patterns, the prognosis of acinar or papillary types is worse than lepidic [
6,
7]. The confirmation of the IAC growth pattern before surgery is essential for the risk stratification of GGN and personalized treatment.
PET/CT has become the primary imaging method for lung cancer evaluation. It can be used to detect and locate the primary tumor, determine the disease stage, or evaluate the treatment effect [
8,
9]. However, whether the preoperative
18F-fluorodeoxyglucose (
18F-FDG) PET/CT can be used to predict the growth pattern of IAC is still unclear [
7,
10,
11]. The maximum standardized uptake value (SUV
max) depends on two factors, the level of glucose uptake, and the spatial distribution of tumor cells. These factors are determined by the growth pattern of each tumor type, which is affected by the proliferation potential of tumor cells. In 2015, Nakamura et al. [
7] first clarified the relationship between SUV
max and individual adenocarcinoma subtypes. The average SUV
max of acinar or papillary types was higher than that of the lepidic type. Son et al. [
10] found that although solid and acinar types showed higher SUV
max since most IACs were lepidic or acinar, there was no significant difference in SUV
max between the main types. Our previous study [
11] also showed similar results as Nakamura et al. Although SUV
max is the only independent factor that can distinguish the growth patterns of IAC, its identification efficacy is still not ideal (AUC = 0.628).
Radiomics is an emerging field in which a large number of objective and quantitative imaging features are explored in order to select the features that are most relevant to clinical, pathological, molecular, and genetic features. This method can increase the accuracy of diagnosis and prognosis and improve treatment efficacy [
12]. The potential of this approach is to quantify the characteristics of tissues or organs beyond the visual interpretation or simple metrics. The texture analysis performed on
18F-FDG PET/CT images has shown great value in diagnosing NSCLC [
13,
14]. In this study, we extracted the texture features of PET and CT images from the respective volume of interest (VOI) and established the PET/CT-based radiomics models to predict intermediate-high risk growth patterns of early IAC.
Discussion
Given the established role of the growth pattern in the early lung adenocarcinoma with GGN, there is a need for non-invasive imaging methods. PET-based SUVmax is a commonly used parameter in the diagnosis of lung cancer. However, it ignores the relationships between two or more voxels, so diagnostic efficiency is not high. In this study, we built a model based on four preoperative radiomic features of 18F-FDG PET/CT images to predict the intermediate-high risk growth pattern in early IAC, and the model showed excellent predictive performance.
The four texture features, including two PET features and two CT features, are all related to image uniformity or heterogeneity. “Sphericity” is a tumor shape descriptor based on PET images, which quantifies the similarity of metabolic tumor volume (MTV) shape and spherical surface. It is entirely defined by the surface of the tumor and therefore only depends on the heterogeneity within the tumor. To a certain extent, segmentation depends on this heterogeneity. Apostolova et al. [
22] studied “asphericity”, the antonym of “sphericity”, and found that asphericity is related to the growth, proliferation, and angiogenesis of NSCLC. Moreover, in adenocarcinoma (ADC), this correlation is much stronger than in squamous cell carcinoma (SCC). In predicting progression-free survival and overall survival, the prognostic power of asphericity is significantly higher than other PET-based parameters (SUV and MTV), clinical and molecular characteristics [
22,
23]. Hyun et al. [
24] used a machine learning algorithm with PET radiomic features to distinguish between ADC and SCC. They found that SCC’s GLZLM_ZLNU is significantly higher than ADC, indicating that SCC is more heterogeneous. Our results also found that sphericity was not easily affected by segmentation methods and quantization levels, which was consistent with the results of Oliver et al. [
25], while GLZLM_ZLNU was also robust to different segmentation methods.
“Kurtosis” derived from the CT histogram reflects the gray distribution in the reaction area. In a practical application, Chae et al. [
26] found that when analyzing GGN, higher kurtosis is a significant difference between preinvasive lesions and IAC. This is consistent with our result that kurtosis of the lepidic group was higher because preinvasive lesions are mainly based on lepidic growth. Besides, Tsubakimoto et al. [
27] found that even in distinguishing ADC and SCC, kurtosis is not as strong as SUV
max, but the diagnostic ability of kurtosis is still strong enough. In the heat map, we found that GLZLM_SZLGE had an excellent negative correlation with HU in conventional indices (especially HUQ1, which represents a low attenuation region; the correlation coefficient was close to − 1). Therefore, it can be considered that CT radiomic features contain the CT
GGO information, so in the end, CT
GGO did not enter the joint model. On HRCT, the GGO component of GGN can indicate a lepidic growth pattern [
28]. The high CT attenuation values of pGGNs suggest IAC [
29], and CT
GGO is an independent predictor of IAC [
30,
31].
We found that the CT signs of the two groups with different IAC growth patterns were mostly overlapped. Among them, the edge was the most promising qualitative CT parameter, and the acinar-papillary group showed a higher proportion of lobulated edge than lepidic group. Lobulation is one of the characteristics of malignant GGN [
32], and it can be used to predict the invasion of GGN [
33]. Moreover, the rad-score that we developed showed a better ability to distinguish the growth patterns. When rad-score was combined with the edge, its clinical value was improved. Besides, we developed a nomogram based on rad-score and edge, which can visualize the prediction results and provide an easy-to-use method for personalized prediction of intermediate-high risk growth patterns.
Our study has some limitations: (1) Although we did internal validation, the single-center design and relatively small sample size may still impair the applicability of the model, especially when it does not include the highest-risk types: solid and micropapillary. Therefore, it is necessary to conduct a standardized multi-center study, expand the sample size, and conduct external validation. (2) This study did not consider the mutation status of EGFR, but the sub-solid nodules have a high EGFR mutation rate [
34]. The subsequent studies should consider EGFR status as a confounding factor. (3) This study has preliminarily demonstrated the potential of radiomics models. In the future, machine learning or deep learning models should be established, in order to improve the predictive performance. (4) The heterogeneity of lung cancer has been shown to play an essential role in disease prognosis [
35]. Due to the short follow-up time, the prognostic value of PET/CT radiomics models for different IAC growth patterns is unclear.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.