Abstract
Objective:
To investigate the effect of reducing pixel size on the consistency of radiomic features and the diagnostic performance of the downstream radiomic signatures for the invasiveness for pulmonary ground-glass nodules (GGNs) on CTs.
Methods:
We retrospectively collected the clinical data of 182 patients with GGNs on high resolution CT (HRCT). The CT images of different pixel sizes (0.8mm, 0.4mm, 0.18 mm) were obtained by reconstructing the single HRCT scan using three combinations of field of view and matrix size. For each pixel size setting, radiomic features were extracted for all GGNs and radiomic signatures for the invasiveness of GGNs were built through two modeling pipelines for comparison.
Results:
The study finally extracted 788 radiomic features. 87% radiomic features demonstrated inter pixel size variation. By either modeling pipeline, the radiomic signature under small pixel size performed significantly better than those under middle or large pixel sizes in predicting the invasiveness of GGNs (p’s value <0.05 by Delong test). With the independent modeling pipeline, the three pixel size bounded radiomic signatures shared almost no common features.
Conclusions:
Reducing pixel size could cause inconsistency in most radiomic features and improve the diagnostic performance of the downstream radiomic signatures. Particularly, super HRCTs with small pixel size resulted in more accurate radiomic signatures for the invasiveness of GGNs.
Advances in knowledge:
The dependence of radiomic features on pixel size will affect the performance of the downstream radiomic signatures. The future radiomic studies should consider this effect of pixel size.
Introduction
Lung cancer is the most common and deadly cancer worldwide and screening high-risk patient using low-dose CT has been showed to reduce lung-cancer mortality. 1 The ground-glass nodule (GGN) on CT is a radiological sign of early stage lung adenocarcinoma and the clinical management relies on the differential diagnosis of its invasiveness during pre-operative assessment. 2 Less-invasive lesions (LIL), including atypical adenomatous hyperplasia (AAH), adenocarcinoma in situ (AIS) and minimally invasive adenocarcinoma (MIA) have favorable prognosis and can be managed by either follow-up or sublobar resection, while invasive adenocarcinoma (IAC) requires lobectomy with poorer prognosis. 3 The studies of characterizing the invasiveness of GGNs on CT showed a trend from identifying semantic features, such as nodule size and solid component, to developing radiomic signature for the invasiveness of GGNs based on quantitative image features of nodule textures, with or without semantic features. 4–7 Radiomic studies achieved prediction accuracy up to 90% in differentiating LIL from IAC. 6,8 These efforts, together with the radiomic studies in other cancer types, 9 demonstrated the potential of radiomics in aiding differential diagnosis of early radiological signs of cancer, such as the GGNs on CT. Nevertheless, these radiomic studies on GGNs did not consider the effect of CT acquisition parameters, such as pixel size and reconstruction algorithms, which might limit the reproducibility of their methods and the generalizability of their conclusion, and thus further investigation is warranted.
The CT acquisition parameters varied greatly in different centers and could affect the consistency of radiomic features as revealed by previous studies. 10–12 For instance, a human study with a small sample of lung cancer patients revealed that nearly 80% of radiomic features had larger intrapatient variability than interpatient variability providing different pixel size. 11 Another human study of lung cancer found that the reconstruction algorithms (standard vs sharp) also influenced the stability of radiomic features. 12 However, the direct utility of radiomics does not reside in radiomic features but in predictive radiomic models or signature that are built upon those features. Few of these reproducibility studies addressed how radiomic models are affected by varying acquisition parameters. Rare studies investigated the effect of acquisition parameters on the diagnostic performance of radiomic models directly. He et al 13 studied the direct link between acquisition parameters and radiomic models for predicting the malignancy of lung nodules and found that CT of no-contrast vs contrast, thin slice vs thick slice, standard reconstruction vs sharp reconstruction resulted in more accurate prediction models. Nevertheless, the study did not make further analysis on whether the model difference is derived from the acquisition-related instability of radiomic features. Therefore, the studies on the effect of acquisition parameters in radiomics are not thorough.
Reconstructed field of view (FOV) and matrix size are two such acquisition parameters that together define the pixel size or spatial resolution of CT images. The reconstruction FOV and matrix size commonly seen in clinics are 400 × 400 mm and 512 × 512, which result in a CT image of 0.8 mm pixel size. Pixel size is essential for the texture analysis as it closely relates to the spatial and intensity information of lesions on CT. 11 Both phantom and human studies showed that the pixel size had a major impact on the consistency of radiomic features. 10,11,14 In these studies, the CT pixel size was above 0.5 mm for human patients and above 0.4 mm for phantoms, which did not cover all possible values encountered in clinical practice. Particularly, high-resolution CT (HRCT) is extensively used in the detection and follow-up for lung nodules. 15 With smaller reconstruction FOV and larger matrix size choices, HRCT can achieve a pixel size as small as 0.18 mm. Moreover, studies found that HRCTs of smaller pixel size provides more visual details on the anatomical structures as rated by radiologists and supposedly facilitate the discrimination of the phenotype of tumor lesions. 16,17 Therefore, HRCT is an ideal choice for studying the links between spatial resolution and radiomics as it enables a wide range of pixel sizes.
In this retrospective study, we collected clinical data of patients with early lung adenocarcinoma, including the pathology diagnosis and HRCT scans, to investigate the effect of pixel size on: (1) the consistency of radiomic features of GGNs and (2) the diagnostic performance of subsequent radiomic signatures for classifying the invasive vs less-invasive nature of the of GGNs.
Methods and materials
Participants
We retrospectively reviewed the medical records of patients with GGNs on HRCT from January 2017 to December 2018 at Shanghai Chest Hospital. The inclusion criteria for this study were: (1) three CT series reconstructed using three combinations of reconstruction FOV and matrix were obtained for the patients’ GGNs. (2) The nodule size (major axis length) of the identified GGN(s) was between 4 and 30 mm. (3) The GGNs were resected and diagnosed as early-stage pulmonary adenocarcinoma by post-operative pathology with histological subtypes, namely AAH, AIS, MIA an IAC. The exclusion criteria are: (1) patients underwent prior invasive procedures for their GGNs, including puncture biopsy and radiofrequency ablation, before admission to our hospital. (2) Patients received cancer treatment for other tumors. (3) The CT image was poor in quality caused by respiratory motion and metal artifact. This study was approved by the institutional review boards of our hospital, and the informed consent from patients was waived for the retrospective nature of this study.
CT imaging acquisition
Pre-operative chest CT was performed with the parameters as follows: X-ray voltage, 120 kVp; tube current, 200 mA; pitch, 0.641; scan FOV, 400 mm. The raw scan data were reconstructed three times using three combinations of reconstruction FOV and matrix size, namely 400 × 400 mm + 512 × 512, 400 × 400 mm + 1024 × 1024 and 180 × 180 mm + 1024 × 1024 (Figure 1). Thus, three CT image series were obtained for each patient and the corresponding pixel sizes were 0.8, 0.4 and 0.18 mm, respectively, which will be later referred as large, middle and small pixel size. In the first two settings, the reconstruction FOV of 400 × 400 mm covered the entire body of the patients while in the third combination the reconstruction FOV of 180 × 180 mm was centered on the patients’ GGNs. Other reconstruction parameters were as follows: reconstruction thickness, 1 mm; reconstruction filter, Sharp; enhancement level, 0; iterative reconstruction, iDose 4 level4.
Figure 1.
Image samples of three CT series reconstructed on a single CT scan of a 40-year-old female patient using three different combinations of reconstruction FOV and matrix size (a–c) resulting in three pixel sizes and providing different spatial resolution for her pulmonary ground-glass nodule (e–f). a represents 400 × 400 mm + 512 × 512 combination, (b) 400 × 400 mm + 1024 × 1024,, (c) 180 × 180 mm + 1024 × 1024. (d) represents the nodule with a pixel size of 0.8 mm, (e) does 0.4 mm and (f) 0.18 mm. FOV, field of view.
Data annotation and preparation
Each GGN was contoured twice separately by two experienced radiologists with 5- and 10 years of experience on a cloud-based annotation platform developed by Deepwise, 1 where radiologists could view DICOM images and annotate them in a nodule-based format (detailed in Supplementary Material 1). The intersection of the two contours was regarded as the final segmentation. Lung window (level −500 Hounsfield unit, HU; width 1500 HU) and mediastinal window (level 40 HU; width 400 HU) were used for displaying the HRCT images during lesion contouring. Meanwhile, each segmented nodule was assigned a specific histological label (AAH, AIS, MIA or IAC) based on the pathological report. As we built binary classification model for the invasiveness of GGNs, AAH, AIS, and MIA were grouped as less-invasive lesions (LIL) and IAC as invasive lesions.
Building radiomic signatures
We extracted 788 radiomic features with 6 categories based on 9 image types for each GGN in 3 pixel sizes respectively using PyRadioimics. 18 Specifically, the nine image types comprised the original image and eight other derived images using wavelet filters. The six categories of features were first-order, shape, gray level co-occurrence matrix (GLCM), gray level dependence matrix (GLDM), gray level run length matrix (GLRLM), gray level size zone matrix (GLSZM).
With these radiomic features of GGNs as input and the pathology-based invasiveness of the same GGNs as output, we constructed logistic regression (LG) model to classify the invasiveness of GGNs. To build comparable classifiers for each pixel size setting, we applied an identical modelling pipeline separately on the three data sets of three distinct pixel sizes. The pipeline consisted of three steps (Figure 2, solid box and arrow), where the first two steps were a sequence of feature selection aiming to filter out the intrapixel-size optimal features most predictive for the invasiveness of GGNs. Firstly, 10 feature candidates were chosen from the 788 features through recursive feature elimination. 19 Then, we fitted a pilot classifier with the 10 feature candidates using all data samples and determined the optimal features whose coefficients were significantly different (p < 0.1) from 0 in the pilot classifier. Finally, the optimal features were employed to build the final classifier as the radiomic signature for the invasiveness of GGNs. For each pixel size setting, this model pipeline resulted in a radiomic signature with its corresponding intrapixel-size optimal features.
Figure 2.
The pipeline of building two versions of radiomic signatures for the three pixel size settings. The intrapixel-size version is represented by the solid black arrows while the cross-pixel-size version is by the dashed brown arrows. The operations, feature selection and fitting classifier, involved in the pipeline are detailed in the methods section. Small, middle and large represent pixel size of 0.18, 0.4 and 0.8 mm, respectively. LR: linear regression. crossOF, cross-pixel-size optimal features; intraOF, intrapixel-size optimal features.
The intrapixel-size optimal features for each pixel size setting might be different, which could hamper the comparability of the corresponding radiomic signatures. We thus took the union of these intrapixel-size optimal features to form the cross-pixel-size optimal features and to build another version of radiomic signature for each pixel size setting (Figure 2, dashed box and arrow). The cross-pixel-size radiomic signatures consisted of exactly the same features and thus more comparable. Both versions of radiomic signatures for each pixel size were trained with fivefold cross-validation. Each signature was identified by three factors, namely the pixel size of the data set it was built on (e.g. Small), the feature composition it involved (e.g. IntraOF standing for intrapixel-size optimal features) and the classifier it used (e.g. LR) (Figure 2).
Statistical analysis
For patients in the two invasiveness groups, Χ2 test was used to evaluate the group difference in categorical variables such as gender and independent t test to evaluate the group difference in continuous variables, such as age and nodule size. To determine the effect of pixel size on the consistency of radiomic features, we conducted one-way repeated measure ANOVA for the 788 radiomic features separately using the data sample of all GGNs in our study cohort. The area under the curve (AUC) of receiver operating characteristic (ROC) was used to evaluate the performance of each radiomic signature. The performance of the radiomic signatures on different pixel sizes were compared using Delong test 20 and the comparison was done separately for the two versions. The modeling and statistical analysis were done using python 3.6 with scikit-learn (v. 0.22), scipy (v. 1.3.0) and researchpy (v. 0.1.8). Two-sided p < 0.05 was considered statistically significant.
Results
Patient characteristics
This study included 182 patients and their clinical characteristics by invasiveness group were summarized in Table 1. In short, the study cohort included 46 males and 136 females with a median age 54 (IQR = 47–62). A total of 197 GGNs were identified with a median nodule size of 12.65 mm (IQR = 9.4–16.68). Patients in two pathology groups were significantly different in terms of gender, age and nodule size (p < 0.05).
Table 1.
Demographic characteristic of the study cohort
| Adenocarcinoma group (N) | Statistics (χ 2 /t test) | ||
|---|---|---|---|
| LIL | IAC | ||
| Sex (N) | 6.15 a | ||
| Males | 24 | 22 | NA |
| Females | 98 | 38 | NA |
| Age, year | |||
| Median (IQR) | 52.0 (46–60) | 56.6 (50–65) | 3.50 c |
| Nodule size (mm) | |||
| Median | 10.79 (10.23–11.35) | 18.54 (17.31–19.66) | 13.57 c |
| Pixel size (mm/pixel) | |||
| Small | 0.173 (0.170–0.177) | 0.174 (0.170–0.178) | 0.303 |
| Middle | 0.400 (0.397–0.403) | 0.401 (0.397–0.405) | 0.455 |
| Large | 0.800 (0.794–0.806) | 0.802 (0.794–0.810) | 0.455 |
AAH, atypical adenomatous hyperplasia; AIS, adenocarcinoma in situ; IAC, invasive adenocarcinoma; IQR, interquartile range; MIA, minimally invasive adenocarcinoma.
p< 0.05.
p<0.01.
p<0.001.
Interpixel-size variation in radiomic features
The one-way repeated measure ANOVA determined that 96.83% radiomic features of the study cohort differed significantly across the three pixel size settings [F (2,390) > 3.05, p < 0.05]. Even when we raised the significance level to p < 0.001 for the ANOVA, the proportion of radiomic features differing significantly across pixel sizes still remained 94.54%. The breakdown of the interpixel-size variation of these radiomic features by image types and feature categories was presented in Figure 3, where over 80% of shape features and majority of first-order and texture features varied by pixel size.
Figure 3.
Heatmap showing the percentage of features differing significantly (p < 0.05) across three pixel sizes among each combination of image type (column) and feature category (row). * Shape feature only applicable for the original image type.
Performance comparison of radiomic signatures
Following the same modeling pipeline, the intrapixel-size radiomic signatures for GGN invasiveness under the three pixel size settings were presented in Table 2. Notably, nearly all the intrapixel-size optimal features selected for each radiomic signature showed significant interpixel-size variations (Table 2, the rightmost column). Meanwhile, the three intrapixel-size signatures almost shared no optimal features and only two shape features (shape_4 MajorAxisLength and shape_2 Flatness) were included in more than one signatures. Based on the results of cross-validation, the performance of radiomic signatures for the three pixel sizes were presented in Table 3. With the intrapixel-size optimal features, the signature under small pixel size (0.18 mm) achieved 0.99 AUC, which was significantly different from both the signature under middle pixel size (0.4 mm) and signature under large pixel size (0.8 mm) (AUCmiddle = 0.96, p < 0.01; AUClarge = 0.96, p < 0.01), as shown in Figure 4a. With the cross-pixel-size optimal features, the difference between small pixel size bounded signature and the other two larger pixel size bounded signatures were still significant (p < 0.01), as shown in Figure 4b.
Table 2.
Feature composition of the three intrapixel-size radiomic signatures for the invasiveness of GGNs
| Radiomic signatures for three pixel size settings | Radiomic features in each radiomic signature | Features’ estimated coefficient in each signature | ANOVA test for feature difference across pixel sizes [F (2,390)] |
|---|---|---|---|
| Small-IntraOF-LR (0.18 mm) |
original_glcm_7 | 1.76 | 420.75 c |
| original_shape_2 | −2.17 | 0.18 | |
| original_shape_4 | 3.72 | 21.47 c | |
| wavelet-HHH_firstorder_7 | 5.12 | 1155.21 c | |
| wavelet-HHL_glrlm_13 | 2.50 | 869.71 c | |
| wavelet-HHL_glszm_5 | 4.75 | 119.17 c | |
| wavelet-LHH_firstorder_7 | −4.61 | 922.70 c | |
| wavelet-LHH_glcm_16 | −3.05 | 242.14 c | |
| wavelet-LLL_glcm_12 | −1.96 | 347.45 c | |
| Middle-IntraOF-LR (0.4 mm) |
original_shape_4 | 3.65 | 21.48 c |
| wavelet-HHL_gldm_9 | 0.83 | 97.58 c | |
| wavelet-LHL_glcm_20 | −2.46 | 1988.33 c | |
| wavelet-LHL_gldm_13 | −1.79 | 246.06 c | |
| wavelet-LHL_glszm_7 | 1.02 | 33.13 c | |
| wavelet-LLL_glcm_13 | −2.29 | 883.30 c | |
| wavelet-LLL_glszm_14 | 1.15 | 2137.55 c | |
| Large-IntraOF-LR (0.8 mm) |
original_shape_2 | −1.06 | 0.18 |
| wavelet-HLH_glcm_12 | 1.90 | 165.45 c | |
| wavelet-HLL_glcm_15 | −1.46 | 332.89 c | |
| wavelet-HLL_glszm_7 | −1.08 | 70.79 c | |
| wavelet-LHL_firstorder_6 | −2.31 | 60.08 c | |
| wavelet-LHL_gldm_5 | 3.24 | 132.16 c | |
| wavelet-LHL_glszm_13 | −2.23 | 182.26 c | |
| wavelet-LLL_glcm_6 | −1.90 | 427.41 c | |
| wavelet-LLL_glcm_7 | 2.16 | 408.02 c |
IntraOF, Intra-pixel-size optimal features; LR, logistic regression.
p < 0.05.
p < 0.01.
p <0.001.
Table 3.
Predictive performance of radiomic signatures built on different pixel size settings in the cross-validation
| Radiomic signatures | Signature score (median, IQR) | AUC (95% CI) | Accuracy | Sensitivity | Specificity | |
|---|---|---|---|---|---|---|
| LIL | IAC | |||||
| Small-IntraOF-LR | 0.01 (0.002–0.08) | 0.99 (0.90–1.00) | 0.986 (0.98–1.00) | 0.95 | 0.97 | 0.92 |
| Middle-IntraOF-LR | 0.07 (0.01–0.25) | 0.92 (0.80–0.99) | 0.958 (0.94–0.98) a | 0.89 | 0.89 | 0.89 |
| Large-IntraOF-LR | 0.04 (0.006–0.26) | 0.93 (0.80–1.00) | 0.958 (0.94–0.98) a | 0.88 | 0.88 | 0.88 |
| Small-CrossOF-LR | 0.01 (0.001–0.06) | 0.99 (0.90–1.00) | 0.985 (0.97–1.00) | 0.94 | 0.96 | 0.92 |
| Middle-CrossOF-LR | 0.11 (0.02–0.25) | 0.93 (0.77–0.99) | 0.934 (0.91–0.96) b | 0.89 | 0.89 | 0.89 |
| Large-CrossOF-LR | 0.05 (0.003–0.22) | 0.93 (0.78–0.99) | 0.949 (0.92–0.97) a | 0.86 | 0.87 | 0.84 |
AAH, Atypical adenomatous hyperplasia; AIS, Adenocarcinoma in situ; CrossOF, Cross-pixel-size optimal features; IAC, Invasive Adenocarcinoma; IQR, interquartile range; IntraOF, Intra-pixel-size optimal features; LR, logistic regression; MIA, minimally invasive adenocarcinoma.
Refer to the Table 2 for the feature composition of each intra pixel-size radiomic signature. The cross-pixel-size radiomic signatures were built with the union of all the intra pixel-size optimal features.
p < 0.01.
p < 0.001 indicating the significance of performance difference between the small-pixel-size based signature and other two pixel-size based signatures.
Figure 4.
ROC curves of radiomic signatures for GGN invasiveness based on different pixel sizes in cross-validation show that Small-LR consistently performs better than Middle-LR and Large-LR with either intra pixel-size optimal features (a) or cross-pixel-size optimal features (b). AUC, area under the curve; CrossOF, Cross-pixel-size optimal features; IntraOF, Intra pixel-size optimal features; ROC, receiver operating characteristic..
Discussion
Radiomics tries to harness the information embedded in the massive medical imaging data, combined with other clinical factors such as pathology and prognosis, to non-less-invasively and automatically characterize tumors. 21 With the promise of assisting patient stratification, treatment decision and therapy response prediction, radiomics has attracted more and more research attentions. 22 However, no consensus exists for the process of building radiomic models and the reproducibility and generalizability of radiomics is still an open question that might have hindered the application of radiomics in clinical practice. 23 Acquisition parameter is one of the major influencing factors and varies greatly in both clinical and research settings. Through building radiomic signatures of the invasiveness of GGNs on HRCTs, this study explored the effect of pixel size determined by reconstructed field of view (FOV) and matrix size on the radiomic features and the performance of the downstream radiomic signatures. We discovered that pixel size of CT images not only causes variation for radiomic features but also results in radiomic signatures performing differently under the same modeling pipeline. The interpixel-size variation of radiomic features found in this study is in line with previous radiomic studies on both phantom and human subjects. 10,11,14 Moreover, we found that this interpixel-size variation could influence the performance of downstream radiomic signatures, specifically for predicting the invasiveness of GGNs on CT. These results together provide further evidence supporting the effort of harmonizing the analytical processes for radiomics. 24
In the study, the radiomic signature built on HRCT of small pixel size (0.18 mm) achieved higher AUC in predicting the invasiveness of GGNs than other two signatures on HRCT of larger pixel size (0.39 mm and 0.78 mm), which indicated that the super high resolution might be helpful in building more accurate radiomic signatures. Previous studies found that super HRCT perceptually provided higher imaging quality for visualizing the microstructure of pulmonary nodules such as nodule edges and the edge of solid component than conventional HRCTs as graded by radiologists. 16,25 As the solid component was critical in making differential diagnosis for the GGNs, 26 HRCT with smaller pixel size, thus improved radiologists’ confidence in predicting the malignancy of GGNs and helped them provide more appropriate follow-up recommendations. 17 Collectively, these findings suggested that higher resolution CTs manifested more discerning information on the phenotype of pulmonary nodules that was not only perceivable to humans but also could be encoded in radiomic features. Contrary to our study where higher resolution of CT resulted in more accurate radiomic signature, however, higher resolution did not help radiologists make better predictions for the pulmonary nodules in Zhu et al 17 ’s study. This discrepancy might be explained by the difference between semantic features and radiomic features in terms of their distinguishing capacities. While semantic features, conceivable to radiologists, were limited in numbers, radiomic features mostly beyond humans’ perception were much larger in quantity, and thus could have captured the extra distinguishing features provided by higher resolution CTs. This explanation was supported by a radiomic study similar to ours showing that semantic feature based model indeed was less accurate than radiomic feature based model in predicting the invasiveness of GGNs. 6
An implication of this study is that to build the most predictive radiomic model of a tumor phenotype requires the availability of multiple imaging data with different properties, for instance different pixel sizes, for the same cohort of patients. In the retrospective scenario, patients meeting inclusive criteria were most likely scanned using very different acquisition parameters and the raw imaging data were usually deleted leaving later reconstruction impossible. Therefore, the radiomic models built in retrospective settings might be suboptimal and should be adopted for clinical tasks that are more error tolerant. In the prospective scenario, it is possible to design scanning and reconstruction protocols in advance to acquire imaging data of multiple properties. Particularly, as image reconstruction does not involve more radiation for patients, one can store multiple imaging data of the same scan. In our center, multiple combinations of FOV and matrix were used to routinely reconstruct CT series for a single scan, which made this study possible. Such routine practice facilitates experiments of model building using different imaging parameters, which are more likely to produce optimal radiomic signatures, and thus is a better strategy for less error tolerant clinical tasks, such as classification of invasiveness of lung nodule that can impact the clinical management.
This study has several limitations. First, the sample size was small and slightly biased in terms of gender, age and nodule size, which could affect the generalizability of the conclusions. Future studies with larger and more balanced data are warranted. Second, pixel size was of three levels but we only quantified the interpixel-size variation of radiomic features between two pixel sizes (small and middle) as this effect for the middle and large pixel sizes was well established in previous studies. Still, it might be more desirable to conduct analysis of variance to evaluate this effect. Third, this study did not explore how other acquisition factors might influence the effect of pixel sizes on radiomics, such as reconstruction thickness, gap, reconstruction filter, iterative reconstruction. Finally, we built the radiomic signature using logistic regression through a customized process and it is unknow whether our findings still hold if different classifiers or modeling pipelines are used.
Conclusion
Pixel size of CT imaging defined by reconstruction FOV and matrix size can cause inconsistency in radiomic features and influence the performance of the subsequent radiomic signatures. HRCT achieved by smaller pixel size may improve the accuracy of radiomic signature in predicting the invasiveness of GGNs.
Footnotes
The authors Guangyu Tao and Lekang Yin contributed equally to the work.
Contributor Information
Guangyu Tao, Email: 121058820@qq.com.
Lekang Yin, Email: lekang2012@163.com.
Dejun Shi, Email: shidejun@deepwise.com.
Jianding Ye, Email: yejianding@126.com.
Zhenghai Lu, Email: lzhlcq@163.com.
Zhen Zhou, Email: zhouzhen@deepwise.com.
Yizhou Yu, Email: yizhouy@acm.org.
Xiaodan Ye, Email: yuanyxd@163.com.
Hong Yu, Email: yuhongchest@163.com.
REFERENCES
- 1. Mazzone PJ, Silvestri GA, Patel S, Kanne JP, Kinsinger LS, Wiener RS, et al. Screening for lung cancer: chest guideline and expert panel report. Chest 2018; 153: 954–85. doi: 10.1016/j.chest.2018.01.016 [DOI] [PubMed] [Google Scholar]
- 2. Goo JM, Park CM, Lee HJ. Ground-glass nodules on chest CT as imaging biomarkers in the management of lung adenocarcinoma. AJR Am J Roentgenol 2011; 196: 533–43. doi: 10.2214/AJR.10.5813 [DOI] [PubMed] [Google Scholar]
- 3. Yanagawa N, Shiono S, Abiko M, Ogata S-ya, Sato T, Tamura G. New IASLC/ATS/ERS classification and invasive tumor size are predictive of disease recurrence in stage I lung adenocarcinoma. J Thorac Oncol 2013; 8: 612–8. doi: 10.1097/JTO.0b013e318287c3eb [DOI] [PubMed] [Google Scholar]
- 4. Fan L, Liu S-Y, Li Q-C, Yu H, Xiao X-S. Multidetector CT features of pulmonary focal ground-glass opacity: differences between benign and malignant. Br J Radiol 2012; 85: 897–904. doi: 10.1259/bjr/33150223 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5. Kim H, Goo JM, Park CM. A simple prediction model using size measures for discrimination of invasive adenocarcinomas among incidental pulmonary subsolid nodules considered for resection. Eur Radiol 2019; 29: 1674–83. doi: 10.1007/s00330-018-5739-x [DOI] [PubMed] [Google Scholar]
- 6. Fan L, Fang M, Li Z, Tu W, Wang S, Chen W, et al. Radiomics signature: a biomarker for the preoperative discrimination of lung invasive adenocarcinoma manifesting as a ground-glass nodule. Eur Radiol 2019; 29: 889–97. doi: 10.1007/s00330-018-5530-z [DOI] [PubMed] [Google Scholar]
- 7. Yu L, Tao G, Zhu L, Wang G, Li Z, Ye J, et al. Prediction of pathologic stage in non-small cell lung cancer using machine learning algorithm based on CT image feature analysis. BMC Cancer 2019; 19: 464. doi: 10.1186/s12885-019-5646-9 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8. Feng B, Chen X, Chen Y, Li Z, Hao Y, Zhang C, et al. Differentiating minimally invasive and invasive adenocarcinomas in patients with solitary sub-solid pulmonary nodules with a radiomics nomogram. Clin Radiol 2019; 74: 570.e1–570.e11. doi: 10.1016/j.crad.2019.03.018 [DOI] [PubMed] [Google Scholar]
- 9. Wang H, Song B, Ye N, Ren J, Sun X, Dai Z, et al. Machine learning-based multiparametric MRI radiomics for predicting the aggressiveness of papillary thyroid carcinoma. Eur J Radiol 2020; 122: 108755. doi: 10.1016/j.ejrad.2019.108755 [DOI] [PubMed] [Google Scholar]
- 10. Shafiq-Ul-Hassan M, Zhang GG, Latifi K, Ullah G, Hunt DC, Balagurunathan Y, et al. Intrinsic dependencies of CT radiomic features on voxel size and number of gray levels. Med Phys 2017; 44: 1050–62. doi: 10.1002/mp.12123 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11. Mackin D, Fave X, Zhang L, Yang J, Jones AK, Ng CS, et al. Harmonizing the pixel size in retrospective computed tomography radiomics studies. PLoS One 2017; 12: e0178524. doi: 10.1371/journal.pone.0178524 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12. Zhao B, Tan Y, Tsai W-Y, Qi J, Xie C, Lu L, et al. Reproducibility of radiomics for deciphering tumor phenotype with imaging. Sci Rep 2016; 6: 23428. doi: 10.1038/srep23428 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13. He L, Huang Y, Ma Z, Liang C, Liang C, Liu Z. Effects of contrast-enhancement, reconstruction slice thickness and convolution kernel on the diagnostic performance of radiomics signature in solitary pulmonary nodule. Sci Rep 2016; 6: 34921. doi: 10.1038/srep34921 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14. Shafiq-ul-Hassan M, Latifi K, Zhang G, Ullah G, Gillies R, Moros E. Voxel size and gray level normalization of CT radiomic features in lung cancer. Sci Rep 2018; 8: 1–9. doi: 10.1038/s41598-018-28895-9 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15. Takahashi S, Tanaka N, Okimoto T, Tanaka T, Ueda K, Matsumoto T, et al. Long term follow-up for small pure ground-glass nodules: implications of determining an optimum follow-up period and high-resolution CT findings to predict the growth of nodules. Jpn J Radiol 2012; 30: 206–17. doi: 10.1007/s11604-011-0033-8 [DOI] [PubMed] [Google Scholar]
- 16. Kakinuma R, Moriyama N, Muramatsu Y, Gomi S, Suzuki M, Nagasawa H, et al. Ultra-high-resolution computed tomography of the lung: image quality of a prototype scanner. PLoS One 2015; 10: e0137165. doi: 10.1371/journal.pone.0137165 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17. Zhu H, Zhang L, Wang Y, Hamal P, You X, Mao H, et al. Improved image quality and diagnostic potential using ultra-high-resolution computed tomography of the lung with small scan FOV: a prospective study. PLoS One 2017; 12: e0172688. doi: 10.1371/journal.pone.0172688 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18. van Griethuysen JJM, Fedorov A, Parmar C, Hosny A, Aucoin N, Narayan V, et al. Computational Radiomics system to decode the radiographic phenotype. Cancer Res 2017; 77: e104–7. doi: 10.1158/0008-5472.CAN-17-0339 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19. Guyon I, Weston J, Barnhill S, Vapnik V. Gene selection for cancer classification using support vector machines. Mach Learn 2002; 46: 389–422. doi: 10.1023/A:1012487302797 [DOI] [Google Scholar]
- 20. Sun X, Xu W. Fast implementation of delong’s algorithm for comparing the areas under correlated receiver operating characteristic curves. IEEE Signal Process Lett 2014; 21: 1389–93. doi: 10.1109/LSP.2014.2337313 [DOI] [Google Scholar]
- 21. Liang Z-G, Tan HQ, Zhang F, Rui Tan LK, Lin L, Lenkowicz J, et al. Comparison of radiomics tools for image analyses and clinical prediction in nasopharyngeal carcinoma. Br J Radiol 2019; 92: 20190271. doi: 10.1259/bjr.20190271 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22. Aerts HJWL, Velazquez ER, Leijenaar RTH, Parmar C, Grossmann P, Carvalho S, et al. Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach. Nat Commun 2014; 5: 4006. doi: 10.1038/ncomms5006 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23. Lambin P, Leijenaar RTH, Deist TM, Peerlings J, de Jong EEC, van Timmeren J, et al. Radiomics: the bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol 2017; 14: 749–62. doi: 10.1038/nrclinonc.2017.141 [DOI] [PubMed] [Google Scholar]
- 24. Zwanenburg A, Leger S, Vallières M, Löck S. Image biomarker standardisation initiative. ArXiv. 2019. Available from: http://arxiv.org/abs/1612.07003.
- 25. Bartlett DJ, Koo CW, Bartholmai BJ, Rajendran K, Weaver JM, Halaweish AF, et al. High-resolution chest computed tomography imaging of the lungs: impact of 1024 matrix reconstruction and photon-counting detector computed tomography. Invest Radiol 2019; 54: 129–37. doi: 10.1097/RLI.0000000000000524 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26. Naidich DP, Bankier AA, MacMahon H, Schaefer-Prokop CM, Pistolesi M, Goo JM, et al. Recommendations for the management of subsolid pulmonary nodules detected at CT: a statement from the fleischner society. Radiology 2013; 266: 304–17. doi: 10.1148/radiol.12120628 [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.