Skip to main content
Download PDF

Evaluation of the invasiveness of pure ground-glass nodules based on dual-head ResNet technique

BMC Cancer volume 24, Article number: 1080 (2024) Cite this article

Abstract

Objective

To intelligently evaluate the invasiveness of pure ground-glass nodules with multiple classifications using deep learning.

Methods

pGGNs in 1136 patients were pathologically confirmed as lung precursor lesions [atypical adenomatous hyperplasia (AAH) and adenocarcinoma in situ (AIS)], minimally invasive adenocarcinoma (MIA), or invasive adenocarcinoma (IAC). Four different models [EfficientNet-b0 2D, dual-head ResNet_3D, a 3D model combining three features (3D_3F), and a 3D model combining 19 features (3D_19F)] were constructed to evaluate the invasiveness of pGGNs using the EfficientNet and ResNet networks. The Obuchowski index was used to evaluate the differences in diagnostic efficiency among the four models.

Results

The patients with pGGNs (360 men, 776 women; mean age, 54.63 ± 12.36 years) included 235 cases of AAH + AIS, 332 cases of MIA, and 569 cases of IAC. In the validation group, the areas under the curve in detecting the invasiveness of pGGNs as a three-category classification (AAH + AIS, MIA, IAC) were 0.8008, 0.8090, 0.8165, and 0.8158 for EfficientNet-b0 2D, dual-head ResNet_3D, 3D_3F, and 3D_19F, respectively, whereas the accuracies were 0.6422, 0.6158, 0.651, and 0.6364, respectively. The Obuchowski index revealed no significant differences in the diagnostic performance of the four models.

Conclusions

The dual-head ResNet_3D_3F model had the highest diagnostic efficiency for evaluating the invasiveness of pGGNs in the four models.

Peer Review reports

Introduction

Lung cancer is the leading cause of cancer-related mortality globally, and the most common subtype is lung adenocarcinoma [1, 2]. The new classification of lung tumors issued by WHO in 2021 classified lung epithelial tumors as lung precursor lesions [atypical adenomatous hyperplasia (AAH) and adenocarcinoma in situ (AIS)], minimally invasive adenocarcinoma (MIA), or invasive adenocarcinoma (IAC) [3]. Studies have confirmed that persistent pure ground-glass nodules (pGGNs) can be pathologically diagnosed as pre-invasive carcinoma or IAC [4, 5], and 30–40% of patients with resected pGGNs are IAC [6, 7]. Therefore, the imaging features of different histopathological subtypes manifesting as pGGNs on computed tomography (CT) overlap significantly.

It is generally known that lung precursor lesions can be followed up without urgent surgical resection. In addition, MIA can be followed up without surgery or treated by partial resection via lung lobectomy. Meanwhile, IAC should be treated with an appropriately extended scope of resection or complete lobectomy if necessary, and lymph node dissection is required [8,9,10]. AIS and MIA have a 5-year disease-free survival rate of 100% [11, 12], whereas that for IAC ranges from 33.3 to 99.0% depending on the major histological subtype [13]. Therefore, it is extremely important to accurately distinguish lung precursor lesions, MIA, and IAC in the new classification to select the best treatment to optimize patient prognosis.

Thus far, studies on the infiltration of pGGNs based on morphological changes on CT have revealed that a larger diameter and mass, higher attenuation value, irregular shape, lobulation, spiculation, cavitation, and internal vascular changes are closely related to the degree of infiltration [5, 7, 14,15,16,17,18,19]. However, it remains difficult to clarify the invasiveness of pGGNs by CT. The possible reasons are as follows: (1) measurements of size and assessments of signs are susceptible to multiple factors (e.g., reproduction of measurement, scanning parameter, reconstruction algorithm, observer experience); (2) some signs have a low incidence in pGGNs and the ability to identify lung precursor lesions is limited; and (3) most features in pre-invasive adenocarcinoma and IAC overlap.

In recent years, increasing numbers of scholars have focused on the application of radiomics for clarifying the invasiveness of pGGNs. Fan et al. [20]. reported that radiomics had good predictive performance in differentiating IAC and non-invasive adenocarcinoma from ground-glass nodules (GGNs). In the external validation group, the area under the curve (AUC) and accuracy (ACC) were 0.936 and 0.881, respectively, suggesting that radiomics could be used as a non-invasive modality for determining follow-up and treatment strategies for lung adenocarcinoma. Other studies demonstrated the good performance of radiomics in predicting IAC with pGGNs and pGGNs invading the pleura, and the AUCs were 0.72 and 0.862, respectively, in the validation group [21, 22]. The radiomic nomogram is helpful for evaluating the invasiveness of pGGNs. It also provides valuable reference information for follow-up and elective surgical treatment. However, radiomic features are derived from manual segmentation, and 3D tumor segmentation is a complex and time-consuming process. At the same time, small blood vessels and bronchi must be avoided during the procedure, but the remaining blood vessels can affect the accuracy of some features. Therefore, both radiological and radiomic models have limitations such as subjective judgment, manual segmentation, and time and labor consumption, which affect the prediction results.

To date, some studies have revealed that artificial intelligence, including convolutional neural networks, is useful for the detection, characterization, and identification of GGNs on CT [23, 24]. Zhao et al. [25] reported that the 3D DenseSharp Network automatically predicted the invasiveness of subcentimeter GGNs (three-category classification) with an ACC of 0.641 and achieved better classification performance than radiologists. Gong et al. [26]. used the deep residual learning network to predict the invasiveness of GGNs (two-category classification: pre-invasive lesions and IAC), and the AUC was 0.92 ± 0.03, reflecting an improved prediction performance for IAC. Lv et al. [27] reported that 3D conventional networks automatically predicted the invasiveness of GGNs (two-category classification), and the AUC was 0.862, similar to that for intraoperative frozen-section analysis. At present, research has used simple 3D deep-learning technology to explore the invasiveness of GGNs.

To the best of our knowledge, no prior study applied a deep learning technique to assess the degree of infiltration of pGGNs. The deep learning network was adopted in the study, which can automatically identify, monitor, qualitative and prognostic judgment of GGNs on CT, with high repeatability, good stability and short time, and can effectively overcome the limitations of subjective judgment, artificial segmentation, time and energy consumption. Therefore, based on EfficientNet-b0 and dual-headed ResNet networks and combinations of different clinical features, four different models were constructed to evaluate the invasiveness of pGGNs and provide a basis for accurate diagnosis and treatment.

Materials and methods

Patient data

In total, 4073 patients with surgically confirmed lung tumors were identified from July 2013 to July 2022. The inclusion criteria were as follows: (1) pathological confirmation of the lesion as IAC, MIA, AIS, or AAH; (2) chest CT revealed pGGNs without solid parts; (3) use of thin-slice chest CT (thickness ≤ 1.5 mm); and (4) no treatment before surgery. The exclusion criteria included a duration of more than 1 month between CT and surgery, incomplete clinical data, and poor CT image quality. Finally, 1136 patients were included in the study. If multiple pGGNs were surgically treated in one patient, one nodule was selected as the study object with infiltration as the main factor and size as the auxiliary factor. Details of the inclusion and exclusion criteria are presented in Fig. 1.

Fig. 1
figure 1

Patient flow chart

Full size image

CT inspection methods

All patients underwent routine CT. The CT scanners were as follows: Somatom Emotion 16 (Siemens, Erlangen, Germany), Definition Flash (Siemens), Force CT (Siemens), Optima CT680 (GE Healthcare, Chicago, IL, USA), and Light Speed 64 (GE Healthcare). The CT scanning parameters were as follows: tube voltage, 100–120 kV; tube current, 200–280 mA; pitch, 0.8–1.0; collimator, 0.6–0.65; and matrix, 512 × 512. The scanning layer and reconstruction thickness were 5 and 1–2 mm, respectively, and standard reconstruction algorithms were used.

Image processing and model construction

In this study, the EfficientNet-b0 and dual-head ResNet networks were used. The entire structure of the 3D model dual-head Res2Net network is presented in Fig. 2. The specific image processing and model construction are presented in the Supplementary file 1.

Fig. 2
figure 2

The entire structure of the 3D model dual-head ResNet network

Full size image

Clinical data and imaging analysis

Clinical data, including sex, age, smoking status, operation status, underlying diseases, tumor markers, and pulmonary diseases, were obtained from the patients’ medical records. CT images were evaluated by two senior physicians with experience in chest diseases (with 12 and 18 respective years of experience), who agreed on the results. CT data were analyzed for the following variables: distribution, lobulation, spiculation, cystic airspace, air bronchogram type, pleural contact type, long diameter (LD), short diameter (SD), maximum CT attenuation (CTmax), minimum CT attenuation (CTmin), mean CT attenuation (CTmean), and standard deviation of CT attenuation (CTsd). In addition, LD and SD were measured for each lesion on the axial image. The CT values were the mean values of multiple measurements taken by two physicians. Pleural contact was divided into three types: direct contact (Type I), pleural tag (Type II), and Type I + Type II (Type III). Four models including the 2D model, 3D model, 3D model combining three clinical features (sex, age, smoking status; 3D_3F model), and 3D model combining 19 clinical features (sex, age, smoking status, operation status, underlying diseases, tumor markers, pulmonary diseases, distribution, lobulation, spiculation, cystic airspace, air bronchogram type, pleural contact type, LD, SD, CTmax, CTmin, CTmean and CTsd).

Statistical analysis

SPSS 25.0 software was used for statistical analysis. Qualitative data were presented as the frequency (component ratio), and the χ2 test was used for intergroup comparisons. Pearson’s correlation analysis was selected when the minimum expected value was greater than 5, and Spearman’s correlation analysis was selected when the minimum expected value was less than or equal to 5 and greater than 1. The Shapiro–Wilk test was used to determine whether the quantitative data were normally distributed. Binary quantitative data with a normal distribution were presented as the mean ± standard deviation, and comparisons between the groups were performed using an independent-samples t-test. Tripartite quantitative data with a normal distribution were presented as the mean median (interquartile range), and comparisons between the groups was performed by ANOVA. The quantitative data of two categories that did not conform to a normal distribution were presented as the median (interquartile range), and the Mann–Whitney U test was used for intergroup comparisons. The quantitative data of three categories that did not conform to a normal distribution were presented as the median (interquartile range), and intergroup comparisons were performed using the Kruskal–Wallis H test (Welch test). The diagnostic performance of artificial intelligence models was evaluatedby the area under the receiver operating characteristic (ROC) curve (AUC), Accuracy(ACC) and other evaluation metrics, such as Recall, Precision, F1-Score. Python software was used to construct the 2D and 3D models and evaluate model efficiency.

Results

General clinical data of the included patients

The clinical data of the 1136 patients are presented in Table 1. The numbers of patients in the training set with AAH + AIS, MIA, and IA were 155, 248, and 392, respectively, whereas the numbers in the validation set were 80, 84, and 177, respectively. Concerning the general clinical data, the mean patient age significantly differed among the groups in both the training and validation sets (both P < 0.05). The incidence of IAC increased with increasing age. In addition, the sex ratio, prevalence of underlying diseases, and tumor marker levels significantly differed among the groups in the training set (all P < 0.05). The distribution of each index was balanced in the training and validation sets.

Table 1 The clinical characteristics of 1136 patients in both training and validation sets
Full size table

CT findings of the study cohort

The CT findings of the 1136 patients are presented in Table 2. Lobulation, spiculation, air bronchogram type, LD, SD, CTmax, CTmin, CTmean, and CTsd significantly differed among the groups in the training and validation sets (all P < 0.05) (Fig. 3). The lesion distribution, cystic airspace, and pleural contact type significantly differed among the groups in the training set (all P < 0.05). The distribution of each index was balanced in the training and validation sets.

Table 2 The CT features of 1136 patients in both training and validation sets
Full size table
Fig. 3
figure 3

CT images showed three pure ground glass nodules (arrows), pathologically representing AIS, MIA and IAC, respectively. Among them, cystic airspace was seen on the MIA nodule, and lobulation on the IAC nodule

Full size image

Comparison of the efficiency of the four models

Figure 4 presents the ROC curves for the four models in the training set. The AUCs for detecting the invasiveness of pGGNs as a three-category classification were 0.7827, 0.8116, 0.8207, and 0.8333 for the EfficientNet-b0 2D, dual-head ResNet_3D, 3D_3F, and 3D_19F models, respectively. Furthermore, the ResNet_3D_3F model has slightly higher diagnostic efficiency than the ResNet_3D_4F and ResNet_3D_6F models (Supplementary Fig.S7 and S8). As presented in Table 3, statistical differences are detected in diagnostic efficiency between the 2D and 3D models using the Obuchowski index (P < 0.05).

Fig. 4
figure 4

ROC curves for the four models in the training set

Full size image
Table 3 The Obuchowski index results of the four models in training set
Full size table

Figure 5 presents the ROC curves of the four models in the validation set. The AUCs for detecting the invasiveness of pGGNs as a three-category classification are 0.80080, 0.80900, 0.81650, and 0.8158 for the EfficientNet-b0 2D, dual-head ResNet_3D, 3D_3F, and 3D_19F models, respectively. Among the four models, the dual-head ResNet_3D_3F model has the highest diagnostic efficiency. Table 4 illustrates that the diagnostic efficiencies of the four models do not differ as tested by the Obuchowski index, but the diagnostic efficiency is higher for the 3D models than for the 2D model.

Fig. 5
figure 5

ROC curves of the four models in the validation set

Full size image
Table 4 The Obuchowski index results of the four models in validation set
Full size table

The confusion matrix for each model is presented in Fig. 6. The figure presents the interval diagram of the distribution of the number of three-category classifications predicted by each model. Table 5 shows the diagnostic efficiencies of the four models in the training set and the validation set. Among the four models, the prediction efficiencies in descending order are Dual-head ResNet_3D_3F, Dual-head ResNet _3D _19F, Dual-head ResNet _3D and Dual-head ResNet_2D, respectively.

Fig. 6
figure 6

The confusion matrix for the four models

Full size image
Table 5 Diagnostic efficiencies of four models in training set and validation set
Full size table

Discussion

To the best of our knowledge, no previous studies used a deep learning technique to specifically evaluate the invasiveness of pGGNs. Therefore, this is the only study to evaluate the invasiveness of pGGNs (three-category classification) using the ResNet network and combining different clinical features. We found that dual-head ResNet_3D_3F model had the highest diagnostic efficiency, whereas there were no significant differences in the diagnostic efficiency compared to other models. In clinical practice, the dual-head ResNet_3D_3F model can be used to evaluate the invasiveness of pGGNs. The results of this study are helpful to intelligently evaluate the invasiveness of pure ground-glass nodules, and provide methodological reference for future related studies.

This study illustrated that the incidence of IAC increases with age, consistent with previous findings [28, 29]. In this group, a mean age exceeding 59 years was useful for categorizing the invasiveness of pGGNs. Some studies revealed that the smoking status is significantly associated with the invasiveness of GGNs [30], contradicting the current findings. It has been reported in the literature that the smoking status is closely related to the growth of lung cancer manifested as GGNs and linked to increased invasiveness [30, 31]. Smokers could be more likely to develop lung adenocarcinoma with subsolid GGNs. In this group, the study population was mainly women, and the overall proportion of smokers was low. These findings could explain the non-significant association of the smoking status with pGGN invasiveness.

Most of the previous literature illustrated that lobulation and spiculation had significance in evaluating the invasiveness of pGGNs [7, 16, 32] and subsolid GGNs [12, 33], in line with the current study findings. However, the utility of the air bronchogram sign in differentiating the invasiveness of GGNs has been inconsistent [29, 33,34,35]. In this group, all patients with air the bronchogram sign had MIA or IAC, and the air bronchogram sign with deformation was more common in IAC. Therefore, we concluded that the air bronchogram sign with deformation was a strong indicator of IAC. Lesion size and density were closely correlated with the invasiveness of GGNs [15, 16, 34, 36]. Hsu et al. [14] found that LD, SD, and the average diameter had significance in identifying the invasiveness of pGGNs, similar to our findings. CT data are important indicators of the invasiveness of pGGNs [16, 37, 38], in line with the present data. CTsd reflects the difference in tumor internal density, and greater values indicate greater tumor heterogeneity, which has not been reported in other studies.

The detection of microscopic changes and hierarchical features invisible to the human eye is one of the advantages of deep learning, and this helps to better distinguish the subtypes of lung cancer [25, 26, 39, 40]. The present study used 2D and 3D models in addition to models combining different clinical features (3 F, 4 F, 6 F, 19 F), differing from the simple 3D deep learning network models reported in other studies [25,26,27]. In this cohort, the ACCs of simple 2D and 3D models for predicting the invasiveness of pGGNs (three-category classification) were 0.6422 and 0.6158, respectively, consistent with the results reported by Zhao [25] and Yu [41]. The AUC and ACC for pGGN invasiveness (three-category classification) evaluated by the dual-head ResNet_3D_3F model were 0.8165 and 0.651, respectively, exceeding the values obtained using other models. We found that precision was more important in the evaluation than the number of clinical features combined. In clinical practice, sex, age, and the smoking status are easy to obtain and simple to implement, highlighting their good prospects for use in future clinical practice. In the dual-head ResNet_3D_3F model, the positive prediction efficiency of lung precursor lesions was the highest (76.9%), followed by IAC (73.7%) and MIA (42.2%). Therefore, patients with lung precursor lesions benefited the most from this model, which can effectively alleviate the anxiety of these patients; Secondly, the benefit was IAC, which required shorter follow-up time or elective surgery.

The ResNet network was used in this study, and it adopted a multi-channel residual connection in the residual block in place of a single-channel residual connection [42]. In this manner, the ResNet network can obtain fine-grained multi-scale features and increase the receptive field of the network. This study differed from prior research by improving multi-scale capability through the use of different resolutions [43, 44]. At the same time, the ResNet module can be inserted into other classic backbone networks such as ResNet [45], ResNeXt [46], and DLA [47] to achieve better results. The dual-head ResNet model adopted in this study first extracts the feature vectors of medical images using the ResNet network and uses the fully connected network to classify the feature vectors, and finally, the classification results obtained by the dual-head model are weighted and combined to obtain the final classification results to obtain better results. Meanwhile, the simple 2D model uses the EfficientNet-B0 network model, the core structure of which is the mobile inverted bottleneck convolution module [48].

Our study had multiple limitations. First, the ResNet network used in this study is relatively simple. It was not compared with other networks, such as ResNet, UNet, and DenseNet, nor was it integrated with other networks. Further research is necessary. Second, external validation was not introduced, which may affect model generalization. Next, we will actively collaborate with more hospitals to conduct external validation to determine the generalization of the model. Finally, the weak explanation of deep learning techniques is also a limitation of this study. The internal mechanisms of deep learning systems should be explored in future studies.

Conclusion

Four different models evaluating pGGNs invasiveness (three-category classification) have good diagnostic efficacy, in which dual-head ResNet_3D_3F model had the highest diagnostic efficiency. In clinical practice, the dual-head ResNet_3D_3F model can be applied to evaluate pGGN invasiveness, thereby providing a valuable tool for accurate diagnosis in patients.

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Abbreviations

AAH:

Atypical adenomatous hyperplasia

AIS:

Adenocarcinoma in situ

MIA:

Minimally invasive adenocarcinoma

IAC:

Invasive adenocarcinoma

CT:

Computed tomography

pGGN:

Pure ground glass nodule

ACC:

Accuracy

AUC:

Area under curve

ResNet:

Residual Network

References

  1. Li M, Narayan V, Gill RR, et al. Computer-aided diagnosis of ground-glass opacity nodules using open-source software for quantifying tumor heterogeneity. AJR Am J Roentgenol. 2017;209(6):1216–27.

    Article  PubMed  PubMed Central  Google Scholar 

  2. Mei XY, Wang R, Yang WJ, et al. Predicting malignancy of pulmonary ground-glass nodules and their invasiveness by random forest. J Thorac Dis. 2018;10(1):458–63.

    Article  PubMed  PubMed Central  Google Scholar 

  3. WHO Classification of Tumours Editorial Board. WHO classification of tumours. Thoracic tumours[M]. 5th ed. Lyon: IARC; 2021.

    Google Scholar 

  4. Chang B, Hwang JH, Choi YH, et al. Natural history of pure ground-glass opacity lung nodules detected by low-dose CT scan. Chest. 2013;143(1):172–8.

    Article  PubMed  Google Scholar 

  5. Wu F, Tian SP, Jin X, et al. CT and histopathologic characteristics of lung adenocarcinoma with pure ground-glass nodules 10 mm or less in diameter. Eur Radiol. 2017;27(10):4037–43.

    Article  PubMed  Google Scholar 

  6. Li Q, Fan L, Cao ET, et al. Quantitative CT analysis of pulmonary pure ground-glass nodule predicts histological invasiveness. Eur J Radiol. 2017;89:67–71.

    Article  PubMed  Google Scholar 

  7. Jin X, Zhao SH, Gao J, et al. CT characteristics and pathological implications of early stage (T1N0M0) lung adenocarcinoma with pure ground-glass opacity. Eur Radiol. 2015;25(9):2532–40.

    Article  PubMed  Google Scholar 

  8. She YL, Zhang L, Zhu HY, et al. The predictive value of CT-based radiomics in differentiating indolent from invasive lung adenocarcinoma in patients with pulmonary nodules. Eur Radiol. 2018;28(12):5121–8.

    Article  PubMed  Google Scholar 

  9. Sim HJ, Choi SH, Chae EJ, et al. Surgical managenment of pulmonary adenocarcinoma presenting as a pure ground-glass nodule. Eur J Cardiothorac Surg. 2014;46(4):632–6.

    Article  PubMed  Google Scholar 

  10. Lee HY, Choi YL, Lee KS, et al. Pure ground-glass opacity neoplastic lung nodules: histopathology, imaging, and management. AJR Am J Roentgenol. 2014;202(3):W224–33.

    Article  PubMed  Google Scholar 

  11. Boland JM, Froemming AT, Wampfler JA, et al. Adenocarcinoma in situ,minimally invasive adenocarcinoma, and invasive pulmonary adenocarcinoma -analysis of interobserver agreement, survival, radiographic characteristics, and gross pathology in 296 nodules. Hum Pathol. 2016;51:41–50.

    Article  PubMed  Google Scholar 

  12. Li M, Wu N, Zhang L, et al. Solid component proportion is an important predictor of tumor, invasiveness in clinical stage T1N0M0 (cT1N0M0) lung adenocarcinoma. Cancer Imaging. 2018;18(1):18.

    Article  PubMed  PubMed Central  Google Scholar 

  13. Murakami S, Ito H, Tsubokawa N, et al. Prognostic value of the new IASLC/ATS/ERS classification of clinical stage IA lung adenocarcinoma. Lung Cancer. 2015;90(2):199–204.

    Article  PubMed  Google Scholar 

  14. Hsu WC, Huang PC, Pan KT, et al. Predictors of invasive adenocarcinomas among pure ground-glass nodules less than 2 cm in Diameter. Cancers (Basel). 2021;13(16):3945.

    Article  PubMed  PubMed Central  Google Scholar 

  15. Chu ZG, Li WJ, Fu BJ, et al. CT characteristics for Predicting Invasiveness in pulmonary pure ground-glass nodules. AJR Am J Roentgenol. 2020;215(2):1–8.

    Article  Google Scholar 

  16. Hu FY, Huang HH, Jiang YY, et al. Discriminating invasive adenocarcinoma among lung pure ground-glass nodules: a multi-parameter prediction model. J Thorac Dis. 2021;13(9):5383–94.

    Article  PubMed  PubMed Central  Google Scholar 

  17. Lim HJ, Ahn S, Lee KS, et al. Persistent pure ground-glass opacity lung nodules ≥ 10 mm in diameter at CT scan: histopathologic comparisons and prognostic implications. Chest. 2013;144(4):1291–9.

    Article  PubMed  Google Scholar 

  18. Son JY, Lee HY, Kim JH, et al. Quantitative CT analysis of pulmonary ground-glass opacity nodules for distinguishing invasive adenocarcinoma from non-invasive or minimally invasive adenocarcinoma:the added value of using iodine mapping. Eur Radiol. 2016;26(1):43–54.

    Article  PubMed  Google Scholar 

  19. Ren H, Liu FF, Xu L, et al. Predicting the histological invasiveness of pulmonary adenocarcinoma manifesting as persistent pure ground-glass nodules by ultra-high-resolution CT target scanning in the lateral or oblique body position. Quant Imaging Med Surg. 2021;11(9):4042–55.

    Article  PubMed  PubMed Central  Google Scholar 

  20. Fan L, Fang MJ, Li ZB, et al. Radiomics signature: a biomarker for the preoperative discrimination of lung invasive adenocarcinoma manifesting as a ground-glass nodule. Eur Radiol. 2019;29(2):889–97.

    Article  PubMed  Google Scholar 

  21. Sun YL, Li C, Jin L, et al. Radiomics for lung adenocarcinoma manifesting as pure ground-glass nodules: invasive prediction. Eur Radiol. 2020;30(7):3650–59.

    Article  PubMed  PubMed Central  Google Scholar 

  22. Jiang YN, Che SY, Ma SC, et al. Radiomic signature based on CT imaging to distinguish invasive adenocarcinoma from minimally invasive adenocarcinoma in pure ground-glass nodules with pleural contact. Cancer Imaging. 2021;21(1):1.

    Article  PubMed  PubMed Central  Google Scholar 

  23. Zhang C, Sun X, Dang K, et al. Toward an Expert Level of Lung Cancer detection and classification using a deep convolutional neural network. Oncologist. 2019;24(9):1159–65.

    Article  PubMed  PubMed Central  Google Scholar 

  24. Huang X, Sun WQ, Tseng TB, et al. Fast and fully-automated detection and segmentation of pulmonary nodules in thoracic CT scans using deep convolutional neural networks. Comput Med Imaging Graph. 2019;74:25–36.

    Article  PubMed  Google Scholar 

  25. Zhao W, Yang JC, Sun YL, et al. 3D deep learning from CT scans predicts Tumor Invasiveness of Subcentimeter Pulmonary Adenocarcinomas. Cancer Res. 2018;78(24):6881–9.

    Article  CAS  PubMed  Google Scholar 

  26. Gong J, Liu JY, Hao W, et al. A deep residual learning network for predicting lung adenocarcinoma manifesting as ground-glass nodule on CT images. Eur Radiol. 2020;30(4):1847–55.

    Article  PubMed  Google Scholar 

  27. Lv YL, Wei Y, Xu K, et al. 3D deep learning versus the current methods for predicting tumor invasiveness of lung adenocarcinoma based on high-resolution computed tomography images. Front Oncol. 2022;21:12:995870.

    Article  Google Scholar 

  28. McWilliams A, Tammemagi MC, Mayo JR, et al. Probability of cancer in pulmonary nodules detected on first screening CT. N Engl J Med. 2013;369(10):910–9.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  29. Zhang T, Zhang CX, Zhong Y, et al. A radiomics nomogram for invasiveness prediction in lung adenocarcinoma manifesting as part-solid nodules with solid components smaller than 6 mm. Front Oncol. 2022;11:12:900049.

    Article  Google Scholar 

  30. Meng FY, Guo Y, Li MY, et al. Radiomics nomogram: a noninvasive tool for preoperative evaluation of the invasiveness of pulmonary adenocarcinomas manifesting as ground-glass nodules. Transl Oncol. 2021;14(1):100936.

    Article  CAS  PubMed  Google Scholar 

  31. Kobayashi K, Sakao K, Deshpande GA, et al. The association between baseline clinical-radiological characteristics and growth of pulmonary nodules with ground-glass opacity. Lung Cancer. 2014;83(1):61–6.

    Article  PubMed  Google Scholar 

  32. Feng H, Shi GF, Xu Q, et al. Radiomics-based analysis of CT imaging for the preoperative prediction of invasiveness in pure ground-glass nodule lung adenocarcinomas. Insights Imaging. 2023;14(1):24.

    Article  PubMed  PubMed Central  Google Scholar 

  33. Yue X, Liu S, Liu S, et al. HRCT morphological characteristics distinguishing minimally invasive pulmonary adenocarcinoma from invasive pulmonary adenocarcinoma appearing as subsolid nodules with a diameter of 3 cm. Clin Radiol. 2018;73(4):e4117–15.

    Article  Google Scholar 

  34. Zhao Z, Xu YN, Yang ZM, et al. Development and validation of a radiomics nomogram for identifying invasiveness of pulmonary adenocarcinomas appearing as subcentimeter ground-glass opacity nodules. Eur J Radiol. 2019;112:161–8.

    Article  PubMed  Google Scholar 

  35. Jiang YN, Xiong ZQ, Zhao WJ, et al. Pathological components and CT imaging analysis of the area adjacent pleura within the pure ground-glass nodules with pleural deformation in invasive lung adenocarcinoma. BMC Cancer. 2022;22(1):958.

    Article  PubMed  PubMed Central  Google Scholar 

  36. Xiang WJ, Xing YF, Jiang S, et al. Morphological factors differentiating between early lung adenocarcinomas appearing as pure ground-glass nodules measuring ≤ 10 mm on thin-section computed tomography. Cancer Imaging. 2014;14(1):33.

    Article  PubMed  PubMed Central  Google Scholar 

  37. Bak SH, Lee HY, Kim JH, et al. Quantitative CT scanning analysis of pure ground-glass opacity nodules predicts further CT scanning change. Chest. 2016;149:180–91.

    Article  PubMed  Google Scholar 

  38. Han L, Zhang P, Wang Y, et al. CT quantitative parameters to predict the invasiveness of lung pure ground-glass nodules (pGGNs). Clin Radiol. 2018;73(5):e5041–7.

    Article  Google Scholar 

  39. Chen LY, Qi H, Lu D, et al. A deep learning based CT image analytics protocol to identify lung adenocarcinoma category and high-risk tumor area. STAR Protoc. 2022;3(3):101485.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  40. Chen LY, Zhang Z. The self-distillation trained multitask dense-attention network for diagnosing lung cancers based on CT scans. Med Phys.2023;16.

  41. Yu Y, Wang N, Huang N, Liu XL, et al. Determining the invasiveness of ground-glass nodules using a 3D multi-task network. Eur Radiol. 2021;31(9):7162–71.

    Article  PubMed  Google Scholar 

  42. Gao SH, Cheng MM, Zhao K, et al. ResNet: a New Multi-scale Backbone Architecture. IEEE Trans Pattern Anal Mach Intel. 2021;43(2):652–62.

    Article  Google Scholar 

  43. Kang E, Min J, Ye JC. A deep convolutional neural network using directional wavelets for low-dose x-ray CT reconstruction. Med Phys. 2017;44(10):e360–75.

    Article  CAS  PubMed  Google Scholar 

  44. Chen H, Zhang Y, Kalra MK, et al. Low-dose CT with a residual encoder-decoder convolutional neural network. IEEE Trans Med Imaging. 2017;36(12):2524–35.

    Article  PubMed  PubMed Central  Google Scholar 

  45. Wang SQ, Wang KK, Yang TP, et al. Improved 3D-ResNet sign language recognition algorithm with enhanced hand features. Sci Rep. 2022;12(1):17812.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  46. Xie S, Girshick R, Dollar P et al. Aggregated residualtr transformations for deep neural networks. IEEE Conf Comput Vis Pattern Recog, pages 5987–95. IEEE.2017.

  47. Ong EJ, Husain S, Bober M. Understanding the distributions of aggregation layers in deep neural networks. IEEE Trans Neural Netw Learn Syst. 2022;5:PP.

    Google Scholar 

  48. Tan MX, Le QV. Efficientnet:Rethinking model scalingfor convolutional neural networks. Proceedings of the 36 th International Conference on Machine Learning (ICML), Long Beach, California. 2019.

Download references

Acknowledgements

We thank Joe Barber Jr., PhD, from Liwen Bianji (Edanz) (www.liwenbianji.cn) for editing the English text of a draft of this manuscript.

Funding

This research was supported by Medical Health Science and Technology Project of Zhejiang Province (Grant Nos. 2021KY602, 2022KY702, and 2023KY1321), Taizhou Science and Technology Plan Project (No.24ywb81) and Hangzhou Medical and Health Science and Technology Project (No. A20220390).

Author information

Authors and Affiliations

  1. Department of Radiology, Taizhou Municipal Hospital, Taizhou, 318000, China

    Dengfa Yang & MinYi Zhao

  2. Department of Radiology, The First Affiliated Hospital of Bengbu Medical University, Bengbu, 233004, China

    Yang Yang

  3. Jianpei Technology, Hangzhou, 311202, China

    Hongli Ji, Bo Hong & Linyang He

  4. Department of Radiology, Sir Run Run Shaw Hospital, Zhejiang University School of Medicine, Hangzhou, 310009, China

    Zhongfeng Niu

  5. Department of Radiology, Anqing Municipal Hospital, Anqing, 246004, China

    Hengfeng Shi

  6. Department of Radiology, Tongde Hospital of Zhejiang Province, Hangzhou, 310012, China

    Meihua Shao & Jian Wang

Authors
  1. Dengfa Yang
    View author publications

    You can also search for this author in PubMed Google Scholar

  2. Yang Yang
    View author publications

    You can also search for this author in PubMed Google Scholar

  3. MinYi Zhao
    View author publications

    You can also search for this author in PubMed Google Scholar

  4. Hongli Ji
    View author publications

    You can also search for this author in PubMed Google Scholar

  5. Zhongfeng Niu
    View author publications

    You can also search for this author in PubMed Google Scholar

  6. Bo Hong
    View author publications

    You can also search for this author in PubMed Google Scholar

  7. Hengfeng Shi
    View author publications

    You can also search for this author in PubMed Google Scholar

  8. Linyang He
    View author publications

    You can also search for this author in PubMed Google Scholar

  9. Meihua Shao
    View author publications

    You can also search for this author in PubMed Google Scholar

  10. Jian Wang
    View author publications

    You can also search for this author in PubMed Google Scholar

Contributions

Guarantor of integrity of the entire study: JW. Study concept and design: DFY, JW. Technical support: HLJ, BH, LYH. Data collection: DFY, YY, MYZ, ZFN, HFS, MHS, JW. Manuscript preparation: DFY, JW. All authors contributed to the article and approved the submitted version.

Corresponding author

Correspondence to Jian Wang.

Ethics declarations

Ethics approval and consent to participate

The experiments involving human participants were reviewed and approved by Institutional Review Board of Taizhou Municipal Hospital. And the requirement of informed consent was waived by Institutional Review Board of Taizhou Municipal Hospital.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.

Additional information

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Supplementary Material 1

Supplementary Material 2

Supplementary Material 3

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Yang, D., Yang, Y., Zhao, M. et al. Evaluation of the invasiveness of pure ground-glass nodules based on dual-head ResNet technique. BMC Cancer 24, 1080 (2024). https://doi.org/10.1186/s12885-024-12823-4

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://doi.org/10.1186/s12885-024-12823-4

Keywords

BMC Cancer

ISSN: 1471-2407

Contact us