Advertisement

A CT-Based Radiomics Nomogram Combined with Clinic-Radiological Characteristics for Preoperative Prediction of the Novel IASLC Grading of Invasive Pulmonary Adenocarcinoma

Open AccessPublished:December 24, 2022DOI:https://doi.org/10.1016/j.acra.2022.12.006

      Rationale and Objectives

      The novel International Association for the Study of Lung Cancer (IASLC) grading system of invasive lung adenocarcinoma (ADC) demonstrated a remarkable prognostic effect and enabled numerous patients to benefit from adjuvant chemotherapy. We sought to build a CT-based nomogram for preoperative prediction of the IASLC grading.

      Materials and Methods

      This work retrospectively analyzed the CT images and clinical data of 303 patients with pathologically confirmed invasive ADC. The histological subtypes and radiological characteristics of the patients were re-evaluated. Radiomics features were extracted, and the optimal subset of features was established by ANOVA, spearman correlation analysis, and the least absolute shrinkage and selection operator (LASSO). Univariate and multivariate analyses identified the independent clinical and radiological variables. Finally, multivariate logistic regression analysis incorporated clinical, radiological, and optimal radiomics features into the nomogram. Receiver operating characteristic (ROC) curve, and accuracy were applied to assess the model's performance. Decision curve analysis (DCA), and calibration curve were applied to assess the clinical usefulness.

      Results

      Nine selected CT image features were used to develop the radiomics model. The accuracy, precision, sensitivity, and specificity of the radiomics model outperformed the clinic-radiological model in the training and testing sets. Integrating Radscore with independent radiological characteristics showed higher prediction performance than clinic-radiological characteristics alone in the training (AUC, 0.915 vs. 0.882; DeLong, p < 0.05) and testing (AUC, 0.838 vs. 0.782; DeLong, p < 0.05) sets. Good calibration and decision curve analysis demonstrated the clinical usefulness of the nomogram.

      Conclusion

      Radiomics features effectively predict high-grade ADC. The combined nomogram may facilitate selecting patients who benefit from adjuvant treatment.

      Key Words

      Abbreviations:

      ADC (Lung adenocarcinoma), CT (Computed tomography), LASSO (Least absolute shrinkage and selection operator), ROC (Receiver operator characteristic), DCA (Decision curve analysis), AUC (Area under the curve), WHO (World Health Organization), IASLC (International Association for the Study of Lung Cancer), MLD (Maximum long-axis diameter), SD (Short-axis diameter), MVD (Maximum vertical diameter), VOI (Volume of interest), ICC (Intra- and interclass correlation coefficients), AIC (Akaike information criterion), OR (Odds ratio), CI (Confidence interval), LOG (Laplacian of Gaussian)

      INTRODUCTION

      According to the latest epidemiological studies, lung cancer is still the leading cause of cancer death worldwide (
      • Siegel RL
      • Miller KD
      • Fuchs HE
      • et al.
      Cancer Statistics, 2021.
      ). Lung adenocarcinoma (ADC) is the most frequent histological subtype of lung cancer and has become the primary cause of mortality from lung cancer (
      • Siegel RL
      • Miller KD
      • Fuchs HE
      • et al.
      Cancer Statistics, 2021.
      ,
      • Cayuela L
      • López-Campos JL
      • Otero R
      The beginning of the trend change in lung cancer mortality trends in Spain, 1980–2018.
      ). Although radical surgical resection, the preferred procedure for patients with early-stage lung cancer, significantly improves prognosis, 30 to 50% of these patients still have recurrence following surgery, which harms their quality of life (

      Abbas AE Surgical management of Lung Cancer: History, Evolution, and Modern advances. 2018 98, DOI 10.1007/s11912-018-0741-7

      ,
      • Uramoto HH
      • Tanaka FF
      Recurrence after surgery in patients with NSCLC.
      ).
      In addition to TNM staging, the histological grading system is vital for assessing the prognosis of various malignancies (
      • Rabe K
      • Snir OL
      • Bossuyt V
      (2019) Interobserver variability in breast carcinoma grading results in prognostic stage differences.
      ,
      • Rice-Stitt T
      • Valencia-Guerrero A
      • Cornejo KM
      Updates in histologic grading of urologic neoplasms.
      ). Although the 2015 World Health Organization (WHO) classification identifies five primary histological subtypes (lepidic, papillary, acinar, solid, and micropapillary) of ADC and proposes grading invasive ADC by major subtypes, the grading method based on major subtypes may underestimate the effect of non-major components, such as a small percentage of the micropapillary and the secondary subtypes, which significantly impact patient prognosis (
      • Bertoglio P
      • Querzoli G
      • Ventura L
      • et al.
      Prognostic impact of lung adenocarcinoma second predominant pattern from a large European database.
      ,
      • Lee G
      • Lee HY
      • Jeong JY
      • et al.
      (2015) Clinical impact of minimal micropapillary pattern in invasive lung adenocarcinoma: prognostic significance and survival outcomes.
      ,
      • Marx A
      • Chan JKC
      • Coindre J
      • et al.
      The 2015 World Health Organization Classification of Tumors of the Thymus: Continuity and Changes.
      ). Furthermore, current research has progressively shown that other complex glandular patterns have the same poor prognosis as the micropapillary and solid subtypes (
      • Zhang R
      • Hu G
      • Qiu J
      • et al.
      Clinical significance of the cribriform pattern in invasive adenocarcinoma of the lung.
      ,
      • Ding Q
      • Chen D
      • Wang X
      • et al.
      Characterization of lung adenocarcinoma with a cribriform component reveals its association with spread through air spaces and poor outcomes.
      ). In fact, several studies have included the addition of pathological characteristics such as secondary histological patterns, nuclear grade, mitotic grade, presence of spread through airspaces (STAS), and necrosis to the primary pattern grading model to improve theprognostic value of the model (
      • Sica G
      • Yoshizawa A
      • Sima CS
      • et al.
      A grading system of lung adenocarcinomas based on histologic pattern is predictive of disease recurrence in stage I tumors.
      ,
      • von der Thüsen JH
      • Tham YS
      • Pattenden H
      • et al.
      Prognostic significance of predominant histologic pattern and nuclear grade in resected adenocarcinoma of the lung: potential parameters for a grading system.
      ,
      • Kadota K
      • Suzuki K
      • Kachala SS
      • et al.
      A grading system combining architectural features and mitotic count predicts recurrence in stage I lung adenocarcinoma. Modern pathology: an official journal of the United States and Canadian Academy of Pathology.
      ,
      • Masai K
      • Sakurai H
      • Sukeda A
      • et al.
      Prognostic impact of Margin Distance and Tumor Spread through Air Spaces in Limited Resection for Primary Lung Cancer.
      ,
      • Liu D
      • Ye Z
      • Chen S
      • et al.
      Novel prognostic model for stratifying survival in stage I lung adenocarcinoma patients.
      ,
      • Bains S
      • Eguchi T
      • Warth A
      • et al.
      (2019) Procedure-specific risk prediction for recurrence in patients undergoing lobectomy or Sublobar Resection for Small (≤2 cm) Lung Adenocarcinoma: An International Cohort Analysis.
      ); nonetheless, most research examined these properties as a single metric and did not take into consideration the heterogeneity of ADC. Considering the limitations of the aforementioned studies, the International Association for the Study of Lung Cancer (IASLC) presented a systematic method for evaluating and incorporating multiple proposed prognostic variables into a grading system (
      • Moreira AL
      • Ocampo PSS
      • Xia Y
      • et al.
      A rading system for invasive Pulmonary Adenocarcinoma: A poposal fom the Int Assocfor the Study of Lung Cancer Pathology Committee.
      ). Their results indicate that the IASLC grading system, based on the primary and high-grade patterns, demonstrated superiority over the previous grading approaches (
      • Moreira AL
      • Ocampo PSS
      • Xia Y
      • et al.
      A rading system for invasive Pulmonary Adenocarcinoma: A poposal fom the Int Assocfor the Study of Lung Cancer Pathology Committee.
      ,
      • Nicholson AG
      • Tsao MS
      • Beasley MB
      • et al.
      The 2021 WHO classification of Lung Tumors: impact of Advances since 2015.
      ). Furthermore, multiple studies have revealed that patients with a predominant high-grade pattern may benefit from adjuvant chemotherapy (
      • Sereno M
      • He Z
      • Smith CR
      • et al.
      Inclusion of multiple high-risk histopathological criteria improves the prediction of adjuvant chemotherapy efficacy in lung adenocarcinoma.
      ,
      • Hung J
      • Yeh Y
      • Jeng W
      • et al.
      Predictive value of the international association for the study of lung cancer/American Thoracic Society/European Respiratory Society classification of lung adenocarcinoma in tumor recurrence and patient survival.
      ), and Deng et al. showed that patients with high-grade ADC in stage Ib-III fared better when treated with adjuvant chemotherapy on the basis of the IASLC grading system (
      • Deng C
      • Zheng Q
      • Zhang Y
      • et al.
      Validation of the Novel International Association for the Study of Lung Cancer Grading System for Invasive Pulmonary Adenocarcinoma and Association With Common Driver Mutations.
      ). Therefore, preoperative identification of different grades of ADC helps predict patient prognosis and select individuals to benefit from adjuvant treatment.
      Current preoperative diagnosis strategies for lung cancer generally involve histological puncture biopsies and CT imaging technology. However, the biopsy only reflects the local status of the tumor, and traditional radiological procedures are subjective qualitative assessments. Therefore, such techniques have inherent restrictions for identifying different ADC grades (
      • Travis WD
      • Brambilla E
      • Noguchi M
      • et al.
      International association for the study of lung cancer/american thoracic society/european respiratory society international multidisciplinary classification of lung adenocarcinoma.
      ,
      • Miao Y
      • Zhang J
      • Zou J
      • et al.
      Correlation in histological subtypes with high resolution computed tomography signatures of early stage lung adenocarcinoma.
      ). Compared to previous approaches, radiomics has once been a hot spot in cancer research, with its non-invasive, repeatable, and quantitative analysis advantages. Radiomics features extracted from medical images of different modalities have the potential to respond to histological differences in tumors (
      • Xv Y
      • Lv F
      • Guo H
      • et al.
      A CT-Based Radiomics Nomogram integrated with Clinic-Radiological Features for Preoperatively Predicting WHO/ISUP Grade of Clear Cell Renal Cell Carcinoma.
      ,
      • Zhang G
      • Xu L
      • Zhao L
      • et al.
      (2020) CT-based radiomics to predict the pathological grade of bladder cancer.
      ,
      • Chen X
      • Zhang Y
      • Chen Y
      • et al.
      MRI-based grading of Clear Cell Renal Cell Carcinoma using a Machine Learning Classifier.
      ). Additionally, many studies have shown that the model's performance may be enhanced by combining radiomics features with other parameters, such as clinical features (
      • Xv Y
      • Lv F
      • Guo H
      • et al.
      A CT-Based Radiomics Nomogram integrated with Clinic-Radiological Features for Preoperatively Predicting WHO/ISUP Grade of Clear Cell Renal Cell Carcinoma.
      ,
      • Peng L
      • Zhang X
      • Liu J
      • et al.
      MRI-radiomics-clinical-based nomogram for prenatal prediction of the placenta accreta spectrum disorders.
      ). However, it is uncertain if radiomics features in conjunction with clinic-radiological characteristics would be used to identify the novel IASLC grade of invasive ADC.
      This study aimed to integrate radiomics, clinical, and radiological features to develop a nomogram for the preoperative grading of invasive ADC.

      MATERIALS AND METHODS

      Patients

      This research included patients who underwent radical surgical resection and had pathological results confirmed as ADC from July 2018 to March 2022 at our institution. A total of 832 patients were initially included (To prevent clustering effects, only the largest lesion among patients who had more than one adenocarcinoma removed was included in the following analysis). Exclusion criteria: (1) adenocarcinoma in situ (n=193), minimally invasive adenocarcinoma (n=236), invasive mucinous adenocarcinoma, and other variants of adenocarcinoma (colloid adenocarcinoma, fetal adenocarcinoma, or enteric-type adenocarcinoma) (n=22); (2) patients who received preoperative treatment (n=5); (3) patients who did not receive a CT scan within one month prior to surgery or severe artifacts on CT images (n=20); (4) lack of clinical or CT data (n=26); (5) unavailable pathologic slides or poor slices quality leading to difficulty to assess histological subtypes (n=27). Finally, 303 individuals with invasive non-mucinous adenocarcinoma were preserved for inclusion in the study (Details of the exclusion criteria and the patient recruitment method were presented in the Appendix Figure A.1). The dataset was randomly allocated in a 7:3 ratio to the training or testing cohorts. All cases in the training cohort were used to train the prediction model, while cases in the testing cohorts were utilized to assess the model's performance independently. The institutional ethical review committee approved the research protocol at our institution. Informed permission was not necessary because of the retrospective nature of the investigation.

      Histopathology Re-Evaluation

      Without clinical information, hematoxylin- and eosin-stained slides of all 303 patients were re-evaluated and graded by two pathologists (F. C. and H. W.) with more than 10 years of experience in pulmonary pathology using a multi-head microscope, and the consensus was reached through discussion. In addition to calculating the five fundamental histological subtype percentages in 5% increments according to the 2015 WHO, non-traditional complex glandular patterns, such as cribriform (defined as nests of neoplastic cells with sieve-like perforations) and fused gland (defined as poorly formed fused glands without intervening stroma or in a ribbon-like formation with irregular borders and single cells infiltrating desmoplastic stroma) patterns, were also evaluated. The specific IASLC grading criteria was showed in the Appendix Table A.1. Herein, due to patients with grade III ADC in the new grading system showing an inferior prognosis and benefit from adjuvant chemotherapy, which is significantly different from patients with grade I and II ADC (
      • Moreira AL
      • Ocampo PSS
      • Xia Y
      • et al.
      A rading system for invasive Pulmonary Adenocarcinoma: A poposal fom the Int Assocfor the Study of Lung Cancer Pathology Committee.
      ,
      • Deng C
      • Zheng Q
      • Zhang Y
      • et al.
      Validation of the Novel International Association for the Study of Lung Cancer Grading System for Invasive Pulmonary Adenocarcinoma and Association With Common Driver Mutations.
      ,
      • Hou L
      • Wang T
      • Chen D
      • et al.
      Prognostic and predictive value of the newly proposed grading system of invasive pulmonary adenocarcinoma in Chinese patients: a retrospective multicohort study. Modern pathology: an official journal of the United States and Canadian Academy of Pathology.
      ,
      • Forest F
      • Laville D
      • Da Cruz V
      • et al.
      WHO grading system for invasive pulmonary lung adenocarcinoma reveals distinct molecular signature: An analysis from the cancer genome atlas database.
      ), all patients were separated into two groups: the low-grade group (grade I and II) and the high-grade group (grade III). The Appendix Figure A.2 displays the five traditional histological subtypes and complex glandular patterns.

      Clinical and Radiological Characteristics

      Demographic information was acquired from the patient's medical records, including gender, age, history of malignancies, family history of cancer, and smoking history. In addition, two radiologists (Z.M.X. with 20 years of work experience and Q.J.H. with 10 years of work experience), who are both thoracic specialists and have extensive expertise in CT diagnosis of lung adenocarcinoma, blinded to histopathological diagnosis, separately documented 13 radiological characteristics comprising lesion involved lobe, nodule shape, nodule type, boundary, lobulation sign, spiculation sign, vacuole sign, air bronchogram sign, vascular convergence sign, pleural retraction sign, MLD (maximum long-axis diameter), SD (short-axis diameter), and MVD (maximum vertical diameter). Any differences were settled by discussion until consensus was established. The interpretation of the different semantic characteristics is presented in the Appendix Table A.2.

      CT Data Acquisition

      All CT plain images were scanned by a 64-slice CT scanner (Discovery CT 750 HD, GE Healthcare, Chicago, IL, USA). All patients were scanned from the apex of the lung to the diaphragm in the supine position with complete inspiration and breath-holding during the chest CT scan. The CT scan parameters were as follows: scan type, helical; rotation time, 0.6s; detector coverage, 40 mm; pitch, 1.375:1; tube voltage, 120 keV; tube current, automatic mA (50–350 mA) for 13 of the noise index; scan field of view (SFOV), Medium Body; scan slice thickness, 5 mm; reconstruction slice thickness and interval, 1.25 mm; reconstruction algorithms, ASIR-V level 40%; reconstruction kernel, stand and lung; matrix size, 512 × 512.

      Image Pre-Processing and Tumor Segmentation

      All CT images were retrieved from the Picture Archiving and Communications System (PACS) in the format of Digital Imaging and Communications in Medicine (DICOM), and the original CT images were properly preprocessed to get the best training results, despite the fact that all images were acquired from the consistent CT scanner. First, the original images were resampled to 1 × 1 × 1 mm3 (x, y, z) to standardize the voxel spacing. Then, a bin width of 25 HU was employed to disperse the voxel intensity and diminish the noise.
      The preprocessed images were loaded into ITK-snap software (version 3.8.0; www.itksnap.org), and reader 1 (Z.A. with 6 years of work experience) performed the volume of interest (VOI) segmentation for all patients by manually sketching the tumor contours on cross-sections of the patient's CT images without knowledge of the pathological results. Two months later, CT scans of 40 patients were randomly selected, and two radiologists (reader 1; reader 2, Z.M.X) separately delineated the tumor VOI to estimate interobserver reproducibility.

      Feature Extraction and Consistency Evaluation

      The radiomics features were extracted using the software of AK (Artificial Intelligence Kit V3.0.0. R, GE Healthcare). Radiomics features were extracted from the original image, wavelet filter, and Laplacian of Gaussian (LOG) filter images. For the LOG filter, the parameters were set to: sigma = 2.0, 3.0; for the wavelet filter, the parameters were set to: level = 1; for the Local Binary Pattern (3D) option, the parameters were set to: level = 2, radius = 1, and subdivision = 1. A total of 1316 radiomics features were extracted from each patient. These features were as follows: (1) first-order statistics (n = 252); (2) shape-based features (n = 14); and (3) texture features (n = 1050), such as the gray level cooccurrence matrix (GLCM), gray level run length matrix (GLRLM), and gray level size zone matrix (GLSZM).
      The intra- and interclass correlation coefficients (ICC) were utilized to evaluate the consistency of features, and features with ICC values higher than 0.8 were considered virtually perfectly consistent (
      • Landis JR
      • Koch GG
      The measurement of observer agreement for categorical data.
      ).

      Radiomics Feature Selection and Signature Construction

      Before analysis, the data were standardized by the Z-score normalization to reduce the bias produced by varying index dimensions. Feature selection includes the following four steps: First, features with an ICC <0.8 were deleted to get stable features. Second, factors that had no significant impact on the grading results were deleted by ANOVA. Third, spearman's correlation analysis was done to lessen the redundancy between features, and if the correlation between features was higher than 0.7, one of them was maintained. Finally, the least absolute shrinkage and selection operator (LASSO) approach with 5-fold cross-validation was utilized to identify the optimal radiomics features. The optimal cutoff value was determined using the Youden index in the ROC analysis of the training dataset. Meanwhile, the Radscore, quantified by a linear combination of each feature and its coefficients, was constructed for each patient. Subsequently, a multivariate logistic regression analysis using the Akaike information criterion (AIC) as the stopping rule was employed to build a radiomics model for predicting high-grade ADC.

      Clinic-Radiological Model Construction

      First, univariate logistic regression analysis chose clinical and radiological parameters with p<0.05 to analyze the differences between high-grade and low-grade ADC. After that, significant risk factors were introduced into the stepwise multivariate logistic regression analysis with the Akaike information criterion (AIC) as the stopping rule to build the clinic-radiological model.

      Combined Model Building and Nomogram Development

      Using multivariate logistic regression analysis, a combined model was built based on Radscore and independent clinical and radiological characteristics; in the meantime, a nomogram was developed to facilitate clinical application.

      Model Comparison and Nomogram Performance Validation

      The performance of the three models was assessed by receiver operator characteristic (ROC) curves, and the area under the curve (AUC) indicate the discriminative performance of the model. The DeLong test was run to compare the ROCs. The accuracy, precision, sensitivity, and specificity of the model were calculated. Calibration curves and the Hosmer-Lemeshow test were employed to assess the apparent performance. The clinical value of the model was tested using decision curve analysis (DCA) by calculating the net benefits at various threshold probabilities in the testing population.

      Statistical Analysis

      All current investigation statistical analyses were done using R 3.5.1 and Python 3.5.6. The R Programming Language was employed to perform the LASSO regression analysis, ROC analysis (with the AUC value calculated), and calibration curve analysis (with the Hosmer-Lemeshow test). The Mann-Whitney U test was employed for the continuous variable, while the Fisher's exact or chi-square test was utilized for the nominal variable. A two-tailed p-value <0.05 indicated statistical significance.

      RESULTS

      Patient Characteristics

      The workflow of this study is shown in Figure 1. A total of 303 patients were included, including 212 patients in the training set (121 in low-grade and 91 in high-grade) and 91 in the testing set (52 in low-grade and 39 in high-grade). There were no statistical differences (p=0.052-0.892) between the training and test sets for all clinical and radiological characteristics (Table 1). Table 1 summarizes the clinical and radiological characteristics of the patients in the training and testing sets. In the training set, there was a significant difference between high-grade and low-grade ADC in terms of nodule type, lobulation sign, spiculation sign, pleural retraction sign, MLD, SD, and MVD (p<0.001).
      Figure 1
      Figure 1The workflow of this study. It is divided into two main modules: radiomics feature engineering and combined model construction and evaluation. First, feature extraction and selection steps established the optimal subset of radiomics features. 1316 radiomics features were extracted from the manually sketched VOI, and feature selection approaches included ANOVA, spearman correlation analysis, and LASSO. Afterward, univariate and multivariate analyses selected independent clinical and radiological characteristics. Finally, a combined model was built by integrating clinical, radiological, and radiomics features through multivariate logistic regression analysis. The nomogram was developed to facilitate clinical application, and the performance of the nomogram was evaluated by DCA, ROC, and calibration curves. VOI, volume of interest; LASSO, least absolute shrinkage and selection operator; DCA, decision curve analysis; ROC, receiver operator characteristic.
      Table 1The Clinical and Radiological Characteristics of Patients in the Training and Testing Sets
      VariableTraining Set (n=212)Testing Set (n=91)p
      Low-gradeHigh-gradepLow-gradeHigh-gradep
      GenderFemale84(69.42%)50(54.95%)0.03136(69.23%)24(61.54%)0.4440.65
      Male37(30.58%)41(45.05%)16(30.77%)15(38.46%)
      Family histories of cancerNever114(94.21%)85(93.41%)0.80849(94.23%)38(97.44%)0.8250.547
      Ever7(5.79%)6(6.59%)3(5.77%)1(2.56%)
      History of malignanciesNever106(87.60%)82(90.11%)0.56948(92.31%)35(89.74%)0.9570.511
      Ever15(12.40%)9(9.89%)4(7.69%)4(10.26%)
      Smoking historyNever112(92.56%)72(79.12%)0.00448(92.31%)35(89.74%)0.9570.276
      Ever9(7.44%)19(20.88%)4(7.69%)4(10.26%)
      Lesion involved lobeRUL33(27.27%)21(23.08%)0.27318(34.62%)15(38.46%)0.9030.313
      RML9(7.44%)8(8.79%)2(3.85%)2(5.13%)
      RLL21(17.36%)27(29.67%)14(26.92%)7(17.95%)
      LUL36(29.75%)23(25.27%)11(21.15%)9(23.08%)
      LLL22(18.18%)12(13.19%)7(13.46%)6(15.38%)
      Nodule shapeRound-or-oval66(54.55%)33(36.26%)0.00821(40.38%)13(33.33%)0.4910.133
      Irregular55(45.45%)58(63.74%)31(59.62%)26(66.67%)
      Nodule typeGround-glass20(16.53%)1(1.10%)<0.0017(13.46%)0(0.00%)<0.0010.814
      Part-solid83(68.60%)26(28.57%)36(69.23%)13(33.33%)
      Solid18(14.88%)64(70.33%)9(17.31%)26(66.67%)
      BoundaryIll-defined55(45.45%)23(25.27%)0.00330(57.69%)7(17.95%)<0.0010.525
      Well-defined66(54.55%)68(74.73%)22(42.31%)32(82.05%)
      Lobulation signNo27(22.31%)3(3.30%)<0.0016(11.54%)0(0.00%)0.0770.062
      Yes94(77.69%)88(96.70%)46(88.46%)39(100.00%)
      Spiculation signNo47(38.84%)2(2.20%)<0.00121(40.38%)3(7.69%)<0.0010.543
      Yes74(61.16%)89(97.80%)31(59.62%)36(92.31%)
      Vacuole signNo92(76.03%)71(78.02%)0.73433(63.46%)29(74.36%)0.270.11
      Yes29(23.97%)20(21.98%)19(36.54%)10(25.64%)
      Air bronchogram signNo66(54.55%)44(48.35%)0.37228(53.85%)21(53.85%)10.754
      Yes55(45.45%)47(51.65%)24(46.15%)18(46.15%)
      Vascular convergence signNo35(28.93%)29(31.87%)0.64417(32.69%)15(38.46%)0.5680.393
      Yes86(71.07%)62(68.13%)35(67.31%)24(61.54%)
      Pleural retraction signNo46(38.02%)12(13.19%)<0.00117(32.69%)7(17.95%)0.1140.86
      Yes75(61.98%)79(86.81%)35(67.31%)32(82.05%)
      Age
      Data are mean ± standard deviation.
      , years
      60.04±10.1661.53±10.900.30859.37±10.0261.95±10.570.2380.892
      MLD
      Data are medians, with interquartile range in parentheses. p is derived from the univariate analysis of each clinical and radiological feature between low- and high-grade invasive ADC patients in the training and testing sets. p* reflects the difference of each clinical and radiological characteristic between the training and testing sets.
      , mm
      12.00(9.00, 17.30)19.00(12.20, 24.00)<0.00116.00(11.00, 22.55)15.00(11.20, 20.00)0.7640.433
      SD
      Data are medians, with interquartile range in parentheses. p is derived from the univariate analysis of each clinical and radiological feature between low- and high-grade invasive ADC patients in the training and testing sets. p* reflects the difference of each clinical and radiological characteristic between the training and testing sets.
      , mm
      12.00(9.00, 16.00)17.00(12.00, 21.00)<0.00115.00(11.45, 20.55)17.00(10.40, 21.80)0.5290.052
      MVD
      Data are medians, with interquartile range in parentheses. p is derived from the univariate analysis of each clinical and radiological feature between low- and high-grade invasive ADC patients in the training and testing sets. p* reflects the difference of each clinical and radiological characteristic between the training and testing sets.
      , mm
      12.00(9.00, 17.00)17.00(12.00, 22.80)<0.00115.00(9.45, 22.00)16.00(11.40, 22.60)0.6910.104
      Abbreviations: ADC, lung adenocarcinoma; LLL, left lower lobe; LUL, left upper lobe; MLD, maximum long-axis diameter; MVD, maximum vertical diameter; RLL, right lower lobe; RML, right middle lobe; RUL, right upper lobe; SD, short-axis diameter;
      Except where indicated, data are numbers of patients, with percentages in parentheses.
      a Data are mean ± standard deviation.
      b Data are medians, with interquartile range in parentheses.p is derived from the univariate analysis of each clinical and radiological feature between low- and high-grade invasive ADC patients in the training and testing sets. p* reflects the difference of each clinical and radiological characteristic between the training and testing sets.

      Clinic-Radiological Model Building

      Univariate analysis indicated that MLD, SD, MVD, nodule type, pleural retraction sign, lobulation sign, and spiculation sign were risk variables for high-grade ADC (Table 2). Multivariate logistic regression analysis indicated that spiculation sign (OR, 8.390; 95%CI, 1.835-38.371), lobulation sign (OR, 7.106; 95%CI, 1.661-30.409), nodule type (OR, 8.399; 95%CI, 4.046-17.436), and MLD (OR, 1.067; 95%CI, 1.015-1.122) remained independent predictors of ADC grade (Appendix Fig. A.3). The clinic-radiological model developed by the four independent predictors showed an AUC of 0.882 (95% CI: 0.842-0.915) in the training set and an AUC of 0.782 (95% CI: 0.703-0.861) in the testing set.
      Table 2Univariate Analysis of Clinical and Radiological Characteristics.
      Univariate Analysis
      OR95%CIp value
      Smoking history3.284(1.408, 7.657)0.006
      MVD, mm1.078(1.037, 1.122)<0.001
      Nodule shape2.109(1.208, 3.683)0.009
      Spiculation sign28.26(6.641, 120.266)<0.001
      Lobulation sign8.425(2.468, 28.758)<0.001
      Gender1.862(1.057, 3.278)0.031
      SD, mm1.082(1.038, 1.129)<0.001
      Boundary2.464(1.362, 4.458)0.003
      Pleural retraction sign4.038(1.986, 8.209)<0.001
      MLD, mm1.09(1.048, 1.134)<0.001
      Nodule type10.643(5.643, 20.072)<0.001
      Abbreviations: ADC, lung adenocarcinoma; MLD, maximum long-axis diameter; MVD, maximum vertical diameter. OR, odds ratio; SD, short-axis diameter.

      Radiomics Features Selection and Model Building

      The 856 characteristics with outstanding stability (ICC > 0.8) were chosen from 1316 features by inter- and intragroup correlation analysis and subsequently subjected to the following feature selection procedure (Fig 1). First, ANOVA identified 400 significant features. Secondly, 49 characteristics were recognized as independent by spearman correlation analysis. Finally, the nine best features were selected by the LASSO regression analysis (Fig 2). The correlation among the nine features was assessed using spearman correlation, and the low correlation among the features revealed that the features had complimentary values (Fig 2c). Subsequently, the nine optimal features were employed to build the radiomics model, which showed a good AUC of 0.897 (95% CI: 0.863-0.928) and accuracy of 0.825 in the training set, and an AUC of 0.819 (95% CI: 0.737-0.894) and accuracy of 0.802 in the testing set, and Fig 3a displays the ROC curve of the testing set. The Appendix Table A.3 describes the details of the nine selected features. The ideal Radscore cutoff value for grade classification was 0.485, based on the maximum Youden index in the training set. Radscore was significantly correlated with ADC grading in both the training set and testing set were significantly correlated (p<0.001) (Fig 3b). The Radscore of high-grade invasive ADC was significantly higher than low-grade in both training and testing sets (Fig 3c and d). The Radscore was calculated as follows Equation 1 :
      Radscore=logsigma20mm3D_gldm_HighGrayLevelEmphasis×0.3226+logsigma20mm3D_glszm_GrayLevelNonUniformity×0.0816+logsigma30mm3D_glcm_InverseVariance×0.1841+logsigma30mm3D_glrlm_ShortRunHighGrayLevelEmphasis×0.195+original_firstorder_Skewness×1.8126+waveletHHH_glszm_LargeAreaEmphasis×0.8448+waveletHLH_gldm_HighGrayLevelEmphasis×0.1939+waveletHLL_glszm_GrayLevelVariance×0.1102+waveletLLH_gldm_DependenceEntropy×0.18160.4589


      Figure 2
      Figure 2LASSO regression in radiomics features engineering. (a), Selection of the optimal tuning parameters by 5-fold cross-validation. Optimal λ value = 1, and log(λ) = 1. (b), LASSO coefficient profiles of radiomics features. The top X-axis represents the number of features, and the bottom X-axis represents the log(λ) values. The blue vertical line at log(λ) value = 1. (c), Spearman rank correlation among the optimal features filtered using LASSO regression. The larger circles and the darker colors imply higher correlations, with red representing negative correlations and green representing positive ones.
      Figure 3
      Figure 3Performance of radiomics features for predicting IASLC grading. (a), ROC curves of radiomics models in the testing set. (b), Radscore comparison between low-grade and high-grade patients. The Radscore significantly differs between high-grade and low-grade patients in the training and testing sets (Wilcoxon, p<0.05). (c), A waterfall plot of the Radscore distribution for all patients in the training set. (d), A waterfall plot of the Radscore distribution for all patients in the testing set.

      Combined Mdel Construction and Model Performance Comparison

      The combined model was constructed using Radscore and the four independent radiological features. The AUCs of the combined model in the training and testing sets were 0.915 (95%CI: 0.885-0.944) and 0.838 (95%CI: 0.764-0.905), respectively. Figure 4a and b compares the performances of the three models in the training and testing sets. Details of the AUC, accuracy, precision, sensitivity, and specificity of each model are presented in Table 3. The discrimination performance of the combined model outperformed the radiomics model in the training set (AUC: 0.915 vs. 0.897; DeLong's test, p < 0.05) and was comparable to the radiomics model in the testing set (AUC: 0.915 vs. 0.897; p = 0.26). The combined model performance outperformed the clinic-radiological model in both the training and testing sets (all p < 0.05).
      Figure 4
      Figure 4Comparison among the ROCs of the clinic-radiological, radiomics, and combined models in the training (a) and testing (b) sets.
      Table 3Comparison of the Prediction Performance Among the Clinic-Radiological, Radiomics, and Combined Models in the Training and Testing Sets.
      ItemCombined ModelRadiomics ModelClinic-Radiological Model
      TrainingTestingTrainingTestingTrainingTesting
      AUC (95%CI)0.915 (0.885, 0.944)0.838 (0.764, 0.905)0.897 (0.863, 0.928)0.819 (0.737, 0.894)0.882 (0.842, 0.915)0.782 (0.703, 0.861)
      Accuracy0.8440.7690.8250.8020.8020.725
      Precision0.8620.7650.80.80.7750.667
      Sensitivity0.7580.6670.7910.7180.7580.718
      Specificity0.9090.8460.8510.8650.8350.731

      Development and Clinical Application of the Nomogram

      Four independent radiological characteristics and the Radscore were applied to generate a nomogram for preoperative prediction of high-grade ADC (Fig 5a). The calibration curve of the nomogram demonstrated good consistency between the prediction probabilities of high-grade ADC and actual observation probabilities in both the training set and testing set (Hosmer-Lemeshow test: p = 0.727 and p = 0.887) (Fig 5b and c). DCA reveals that the nomogram has a higher net benefit than the clinic-radiological model in predicting high-grade ADC in the testing set (Fig 5d). Figure 6 displays high-grade and low-grade cases identified by the nomogram, and Fig 6g illustrates the difference between high-grade and low-grade ADC cases for nine radiomics features.
      Figure 5
      Figure 5Performance and clinical application of the nomogram. (a), The nomogram was generated by merging Radscore and four clinic-radiological features in the training set. Calibration curves revealed satisfactory calibration of the nomogram in the training (b) and testing (c) sets. (d) DCA of the prediction model in the testing set. The Y-axis shows the net benefit, and the x-axis represents the threshold probability. The DCA demonstrates that the nomogram provides higher net benefits than the clinic-radiological model in predicting high-grade invasive ADC in the overwhelming majority of areas. ADC, lung adenocarcinoma; MLD, maximum long-axis diameter (1, ground-glass; 2, part-solid; 3, solid).
      Figure 6
      Figure 6The comparison of patients with high- and low-grade invasive ADCs. a–c, Patient 1: A 69-year-old female patient with low-grade ADC, including a histological subtype of 40% lepidic (b) and 60% acinar (c) patterns (× 40). The CT image shows a part-solid nodule with lobulation, pleural retraction, vacuole, and vascular convergence signs (arrows). (d–f), Patient 2: A 53-year-old female patient with high-grade ADC suffering from histological subtypes accounting for 80% solid (e) and 20% complex glandular (f) patterns (× 40). (d), The CT image exhibits a solid nodule with lobulation, short spiculation, and vascular convergence signs (arrows). (g), Comparison of 9 radiomics features of the two patients.

      DISCUSSION

      In this work, we sought to develop optimal radiomics features to preoperative prediction of the IASLC grading of invasive ADC, and the final results showed that the accuracy, precision, and specificity of the radiomics model outperformed the clinic-radiological model both in training and testing sets. Furthermore, good calibration and decision curve analysis demonstrated the clinical usefulness of the nomogram.
      Increasing evidence suggests that histological subtypes associated with poor prognosis in ADC should be considered with complex glandular patterns(
      • Zhang R
      • Hu G
      • Qiu J
      • et al.
      Clinical significance of the cribriform pattern in invasive adenocarcinoma of the lung.
      ,
      • Ding Q
      • Chen D
      • Wang X
      • et al.
      Characterization of lung adenocarcinoma with a cribriform component reveals its association with spread through air spaces and poor outcomes.
      ). Moreover, the concepts for micropapillary adenocarcinoma of the lung should be broadened to include the filigree and the classical pattern since both have poor prognoses (
      • Emoto K
      • Eguchi T
      • Tan KS
      • et al.
      Expansion of the Concept of Micropapillary Adenocarcinoma to include a Newly Recognized Filigree pattern as well as the Classical Pattern Based on 1468 Stage I Lung Adenocarcinomas.
      ). A lack of awareness of these structures will lead to the misclassification of prognostic outcomes. Another noteworthy point is that the ratio of various subtypes impacts the prognosis of individuals with ADC (
      • Lee G
      • Lee HY
      • Jeong JY
      • et al.
      (2015) Clinical impact of minimal micropapillary pattern in invasive lung adenocarcinoma: prognostic significance and survival outcomes.
      ,
      • Bertoglio P
      • Querzoli G
      • Ventura L
      • et al.
      Prognostic impact of lung adenocarcinoma second predominant pattern from a large European database.
      ). The IASLC grading system, based on the primary pattern versus the high-grade pattern, demonstrated superiority over the previous grading approaches (
      • Moreira AL
      • Ocampo PSS
      • Xia Y
      • et al.
      A gading system for Invasive Pulmonary Adenocarcinoma: aproposal from the International Association for the Study of Lung Cancer Pathology Committee.
      ). In addition, Fujikawa, Deng, et al. indicated genotypic variations across different grades of ADC, with KRAS mutations and AKL fusions predominating in patients with high-grade ADC, whereas EGFR mutations were more prevalent in low-grade ADC (
      • Rokutan-Kurata M
      • Yoshizawa A
      • Ueno K
      • et al.
      Validation Study of the International Association for the Study of Lung Cancer Histologic Grading System of Invasive Lung Adenocarcinoma.
      ,
      • Fujikawa R
      • Muraoka Y
      • Kashima J
      Clinicopathologic and Genotypic Features of Lung Adenocarcinoma Characterized by the International Association for the Study of Lung Cancer Grading System.
      ). Deng et al. also showed that stage Ib-III patients with high-grade ADC fared better when treated with adjuvant chemotherapy, suggesting the potential for more extensive treatment with postoperative chemotherapy regimens in this subgroup (
      • Deng C
      • Zheng Q
      • Zhang Y
      • et al.
      Validation of the Novel International Association for the Study of Lung Cancer Grading System for Invasive Pulmonary Adenocarcinoma and Association With Common Driver Mutations.
      ). In brief, preoperative grading of ADC not only predicts patient prognosis but also enables the study of possible genetic variations in patients and the formulation of the treatment approach. However, preoperative histological grading is hampered by puncture biopsy's random and invasive nature and traditional radiology methods' subjective and qualitative nature. Our work obtained high-throughput radiomics features to identify the histological heterogeneity of ADC from the entire tumor area of the patient's CT images with trustworthy results.
      Our study indicates that CT-based radiomics features may describe tumor histological heterogeneity, which is analogous to many earlier studies (
      • Zhang G
      • Xu L
      • Zhao L
      • et al.
      (2020) CT-based radiomics to predict the pathological grade of bladder cancer.
      ,
      • Chen X
      • Zhang Y
      • Chen Y
      • et al.
      MRI-based grading of Clear Cell Renal Cell Carcinoma using a Machine Learning Classifier.
      ). Several published radiomics studies have explored the relevance of the preoperative histological subtype categorization of ADC (
      • Song SH
      • Park H
      • Lee G
      • et al.
      Imaging Phenotyping using Radiomics to Predict Micropapillary Pattern within Lung Adenocarcinoma.
      ,
      • Park S
      • Lee SM
      • Noh HN
      • et al.
      (2020) Differentiation of predominant subtypes of lung adenocarcinoma using a quantitative radiomics approach on CT.
      ,
      • Chen L
      • Yang S
      • Wang H
      • et al.
      Prediction of micropapillary and solid pattern in lung adenocarcinoma using radiomic values extracted from near-pure histopathological subtypes.
      ,
      • Bae JM
      • Jeong JY
      • Lee HY
      • et al.
      Pathologic stratification of operable lung adenocarcinoma using radiomics features extracted from dual energy CT images.
      ,
      • Xu Y
      • Ji W
      • Hou L
      • et al.
      Enhanced CT-Based Radiomics to predict Micropapillary pattern within Lung Invasive Adenocarcinoma.
      ). Song, Xu et al. extracted radiomics characteristics from CT images to predict the existence of ADC subtypes; nevertheless, all of these investigations predicted only the presence of micropapillary while disregarding other high-risk subtypes (
      • Song SH
      • Park H
      • Lee G
      • et al.
      Imaging Phenotyping using Radiomics to Predict Micropapillary Pattern within Lung Adenocarcinoma.
      ,
      • Xu Y
      • Ji W
      • Hou L
      • et al.
      Enhanced CT-Based Radiomics to predict Micropapillary pattern within Lung Invasive Adenocarcinoma.
      ). Yang et al. developed an algorithm for micropapillary and solid ADC subtypes identification utilizing radiomics features derived from near pure subtypes (>70 %) of ADC and patch-wise image analysis, with accuracies of 91.6 % (87/95) to 94.4 % (34/36)((
      • Chen L
      • Yang S
      • Wang H
      • et al.
      Prediction of micropapillary and solid pattern in lung adenocarcinoma using radiomic values extracted from near-pure histopathological subtypes.
      ). However, the fact that most ADC histological subtypes are mixed would hamper the practical implementation of the “near pure” system. Park, Bae, et al. revealed that CT-based radiomics features might assist in predicting the histologic classification of ADC (
      • Park S
      • Lee SM
      • Noh HN
      • et al.
      (2020) Differentiation of predominant subtypes of lung adenocarcinoma using a quantitative radiomics approach on CT.
      ,
      • Bae JM
      • Jeong JY
      • Lee HY
      • et al.
      Pathologic stratification of operable lung adenocarcinoma using radiomics features extracted from dual energy CT images.
      ). However, these investigations were based on the primary subtype grading approach. So many individuals categorized as grade III by the IASLC grading approach would be classified as grade II by the primary subtype approach. Our research seeks to identify these individuals from those with primarily glandular or papillary subtypes since the two have entirely different prognoses (
      • Rokutan-Kurata M
      • Yoshizawa A
      • Ueno K
      • et al.
      Validation Study of the International Association for the Study of Lung Cancer Histologic Grading System of Invasive Lung Adenocarcinoma.
      ).
      The combined model was constructed from Radscore and four radiological features. For the fused radiomics signatures, nine features were eventually chosen by LASSO, comprising one first-order feature, four Laplacian of Gaussian (LOG) texture features, and four wavelet texture features. The first-order features define the distribution pattern of grey-level pixel values inside a complete tumor. Furthermore, the LOG filter is an edge-detecting filter that uses Gaussian smoothing to eliminate noise, and the wavelet filter reveals spatial variability across multiple dimensions by splitting the original picture in various directions. Our results indicate that the two features HHH_glszm_LargeAreaEmphasis based on the wavelet transform and the original first_order_Skewness display the strongest correlation with ADC grading. Skewness quantifies the asymmetry of the distribution of values around the mean value. Moreover, HHH_glszm_LargeAreaEmphasis is a measure of the distribution of large area size zones, with a more excellent value indicative of larger size zones and more coarse textures. Several studies have underlined the advantages of filtering image features for tumor subtype classification (
      • Lubner MG
      • Stabo N
      • Abel EJ
      • et al.
      CT Textural analysis of Large Primary Renal Cell Carcinomas: pretreatment Tumor Heterogeneity correlates with Histologic Findings and Clinical outcomes.
      ,
      • Skogen K
      • Ganeshan B
      • Good C
      • et al.
      Measurements of heterogeneity in gliomas on computed tomography relationship to tumour grade.
      ,
      • Ganeshan B
      • Goh V
      • Mandeville HC
      • et al.
      Non-small cell lung cancer: histopathologic correlates for texture parameters at CT.
      ). Finally, considering the value of clinical and radiological characteristics in diagnosing ADC, we screened four radiological characteristics by univariate and multivariate logistic regression analysis. We merged them with radiomics features to construct a combined model. However, the AUCs of the combined nomogram achieved a slight increase over the radiomics model in both the training (from 0.897 to 0.915) and testing (from 0.819 to 0.838) sets, indicating that the radiological characteristics provide valuable but finite complementary information. Moreover, most of the high-grade ADC presented as solid on CT imaging, which is consistent with the results of Fujikawa et al. (
      • Fujikawa R
      • Muraoka Y
      • Kashima J
      Clinicopathologic and Genotypic Features of Lung Adenocarcinoma Characterized by the International Association for the Study of Lung Cancer Grading System.
      ). The DCA analysis demonstrated that the final developed nomogram provides an essential reference for physicians to construct tailored treatment regimens and follow-up care.
      We recognize several limitations in our work. First, the patient data for this retrospective analysis were gathered at a single institution, which may have contributed to selection bias. Second, since the IASLC grading system goes exclusively to invasive non-mucinous adenocarcinoma, our research did not include the uncommon types of invasive adenocarcinoma (invasive mucinous adenocarcinoma, and other variants of adenocarcinoma). Nevertheless, a model incorporating all types of invasive ADC may be more sufficient for practical clinical work. Third, the CT images were acquired from a single scanner, and radiomics shows significant intra- and inter-scanner heterogeneity; therefore, the generalization of the model needs to be further validated. Finally, our investigation described the tumor histological differences exclusively from the unimodality data; integrating radiomics data with other multiple scales and modalities is an emerging discipline, enabling a more specific and unambiguous biological interpretation of cancer histological heterogeneity and prognostic treatment outcomes (
      • Lu C
      • Shiradkar R
      • Liu Z
      Integrating pathomics with radiomics and genomics for cancer prognosis: abrief review.
      ). Our subsequent investigation will explore the correlation between radiomics, pathomics, and genomics in ADC to offer a better biological explanation for the patient.
      To summarize, radiomics features effectively predict high-grade ADC. The combined nomogram may facilitate selecting patients who benefit from adjuvant treatment; however, more extensive prospective studies are necessary to further validate radiomics features' ability to predict high-grade ADC.

      Funding

      This research was supported by grants from the National Natural Science Foundation of China [No. 82171931], the Science and Technology Program of Guangzhou [Nos. 201903010032 and 202102080572], and the Panyu Science and Technology Program of Guangzhou [Nos. 2019-Z04-01, 2019-Z04-23, and 2022-Z04-013].

      Appendices

      Figures A.1A.3 and Tables A.1A.3.
      Figure A2
      Figure A.2The five traditional subtypes of lung adenocarcinoma defined by the 2015 WHO: (a) lepidic, (b) acinar, (c) papillary, (d) micropapillary, and (e) solid; Non-traditional subtype: (f) complex glandular
      Figure A3
      Figure A.3Forest plot of the multivariate analysis. OR, odds ratio; MLD, maximum long-axis diameter.
      Appendix Table A.1The Novel IASLC Grading System of Invasive Lung Adenocarcinoma
      GradePercentage of Subtypes
      Grade Ilepidic subtype with <20% of the high-grade pattern (solid, micropapillary, and complex glandular patterns)
      Grade IIpapillary or acinar subtype with <20% of high-grade pattern
      Grade IIIany predominant subtype with ≥20% of high-grade pattern
      Appendix Table A.2The Definition of Radiological Characteristics
      Radiological featuresDescription
      lesion involved lobeThe location of the tumor in the lung. Classified as RUL, RML, RLL, LUL, LLL.
      Nodule shapeSplit into round-or-oval and irregular.
      Nodule typeCategorized as: ground-glass, part-solid and solid [
      • Naidich DP
      • Bankier AA
      • MacMahon H
      • et al.
      Recommendations for the management of subsolid pulmonary nodules detected at CT: a statement from the Fleischner Society.
      ,
      • Hansell DM
      • Bankier AA
      • MacMahon H
      • et al.
      Fleischner Society: glossary of terms for thoracic imaging.
      ].
      BoundaryThe tumor boundaries are well-defined or ill-defined.
      Lobulation signThe margins of the tumor present multiple curvilinear bumps
      • Hansell DM
      • Bankier AA
      • MacMahon H
      • et al.
      Fleischner Society: glossary of terms for thoracic imaging.
      .
      Spiculation signThe short, unbranched linear shadow not attached to the pleura extends peripherally from the lesion area

      Li F, Sone S, Abe H, et al. Malignant versus benign nodules at CT screening for lung cancer: comparison of thin-section CT findings. 2004 793–798..

      . It is indicated by its existence (yes) and absence (no).
      Air bronchogram signInflatable bronchioles are visible within the tumor. It is indicated by its existence (yes) and absence (no).
      vascular convergence signMultiple vessels converge toward the tumor. It is indicated by its existence (yes) and absence (no).
      Pleural retraction signThe tumor is attached to the adjacent pleura
      • Lee E
      • Son D
      • Kim S
      • et al.
      Prediction of recurrence-free survival in postoperative non-small cell lung cancer patients by using an integrated model of clinical information and gene expression.
      . It is indicated by its existence (yes) and absence (no).
      MLDThe longest diameter of the tumor in the largest cross-section.
      SDThe shortest diameter of the tumor in the largest cross-section.
      MVDThe maximum thickness of the tumor.
      Vacuole signAir-containing hypodense areas of 1-3 mm within the lesion. It is indicated by its existence (yes) and absence (no).
      Abbreviations: LLL, left lower lobe; LUL, left upper lobe; MLD, maximum long-axis diameter; MVD, maximum vertical diameter RLL, right lower lobe; RML, right middle lobe; RUL, right upper lobe; SD, short-axis diameter..
      Appendix Table A.3Selected Features and Corresponding Coefficients
      VariablesFilterFeature ClassCoefficients95%CI
      High Gray Level EmphasisLOG (δ=2)GLDM0.3226-0.1689, 0.814
      Gray Level Non UniformityLOG (δ=2)GLSZM0.0816-0.6313, 0.7944
      Inverse VarianceLOG (δ=3)GLCM-0.1841-0.7091, 0.341
      Short Run High Gray Level EmphasisLOG (δ=3)GLRLM0.195-0.3065, 0.6965
      SkewnessNoneFirst order-1.8126-2.5858, -1.0393
      Large aea emphasisWavelet (HHH)GLSZM0.84480.1534, 1.5361
      High gay level emphasisWavelet (HLH)GLDM0.1939-0.3484, 0.7361
      Gray level varianceWavelet (HLL)GLSZM0.1102-0.3974, 0.6178
      Dependence entropyWavelet (LLH)GLDM0.1816-0.3256, 0.6888
      Abbreviations: GLCM, gray level cooccurrence matrix; GLDM, gray level dependence matrix; GLRLM, gray level run length matrix; GLSZM, gray level size zone matrix;LOG, laplacian of Gaussian;
      The value in brackets indicate the filters (L: low-pass filter, H: high-pass filter) applied in the x, y and z directions.

      References

        • Siegel RL
        • Miller KD
        • Fuchs HE
        • et al.
        Cancer Statistics, 2021.
        CA: a cancer j clinic. 2021; 71: 7-33https://doi.org/10.3322/caac.21654
        • Cayuela L
        • López-Campos JL
        • Otero R
        The beginning of the trend change in lung cancer mortality trends in Spain, 1980–2018.
        Archivos de Bronconeumología. 2021; 57: 115-121https://doi.org/10.1016/j.arbres.2020.04.025
      1. Abbas AE Surgical management of Lung Cancer: History, Evolution, and Modern advances. 2018 98, DOI 10.1007/s11912-018-0741-7

        • Uramoto HH
        • Tanaka FF
        Recurrence after surgery in patients with NSCLC.
        Translatlung cancer res. 2014; 3: 242-249https://doi.org/10.3978/j.issn.2218-6751.2013.12.05
        • Rabe K
        • Snir OL
        • Bossuyt V
        (2019) Interobserver variability in breast carcinoma grading results in prognostic stage differences.
        Hum Pathol. 2019; 94: 51-57https://doi.org/10.1016/j.humpath.2019.09.006
        • Rice-Stitt T
        • Valencia-Guerrero A
        • Cornejo KM
        Updates in histologic grading of urologic neoplasms.
        Arch Pathol Lab Med. 2020; 144: 335-343https://doi.org/10.5858/arpa.2019-0551-RA
        • Bertoglio P
        • Querzoli G
        • Ventura L
        • et al.
        Prognostic impact of lung adenocarcinoma second predominant pattern from a large European database.
        J surg oncol. 2021; 123: 560-569https://doi.org/10.1002/jso.26292
        • Lee G
        • Lee HY
        • Jeong JY
        • et al.
        (2015) Clinical impact of minimal micropapillary pattern in invasive lung adenocarcinoma: prognostic significance and survival outcomes.
        Am j surgi pathol. 2015; 39: 660-666https://doi.org/10.1097/PAS.0000000000000399
        • Marx A
        • Chan JKC
        • Coindre J
        • et al.
        The 2015 World Health Organization Classification of Tumors of the Thymus: Continuity and Changes.
        J thoracic oncol: official pub Intl Assoc Study Lung Cancer. 2015; 10: 1383-1395https://doi.org/10.1097/JTO.0000000000000654
        • Zhang R
        • Hu G
        • Qiu J
        • et al.
        Clinical significance of the cribriform pattern in invasive adenocarcinoma of the lung.
        J Clinic Pathol. 2019; 72: 682-688https://doi.org/10.1136/jclinpath-2019-205883
        • Ding Q
        • Chen D
        • Wang X
        • et al.
        Characterization of lung adenocarcinoma with a cribriform component reveals its association with spread through air spaces and poor outcomes.
        Lung Cancer. 2019; 134: 238-244https://doi.org/10.1016/j.lungcan.2019.06.027
        • Sica G
        • Yoshizawa A
        • Sima CS
        • et al.
        A grading system of lung adenocarcinomas based on histologic pattern is predictive of disease recurrence in stage I tumors.
        The American journal of surgical pathology. 2010; 34: 1155-1162https://doi.org/10.1097/PAS.0b013e3181e4ee32
        • von der Thüsen JH
        • Tham YS
        • Pattenden H
        • et al.
        Prognostic significance of predominant histologic pattern and nuclear grade in resected adenocarcinoma of the lung: potential parameters for a grading system.
        Jthoracic oncol: official pub Intl Assoc Study Lung Cancer. 2013; 8: 37-44https://doi.org/10.1097/JTO.0b013e318276274e
        • Kadota K
        • Suzuki K
        • Kachala SS
        • et al.
        A grading system combining architectural features and mitotic count predicts recurrence in stage I lung adenocarcinoma. Modern pathology: an official journal of the United States and Canadian Academy of Pathology.
        Inc. 2012; 25: 1117-1127https://doi.org/10.1038/modpathol.2012.58
        • Masai K
        • Sakurai H
        • Sukeda A
        • et al.
        Prognostic impact of Margin Distance and Tumor Spread through Air Spaces in Limited Resection for Primary Lung Cancer.
        J thoracic oncol: official pub Intl Assoc Study Lung Cancer. 2017; 12: 1788-1797https://doi.org/10.1016/j.jtho.2017.08.015
        • Liu D
        • Ye Z
        • Chen S
        • et al.
        Novel prognostic model for stratifying survival in stage I lung adenocarcinoma patients.
        Jcancer res clinic oncol. 2020; 146: 801-807https://doi.org/10.1007/s00432-019-03110-y
        • Bains S
        • Eguchi T
        • Warth A
        • et al.
        (2019) Procedure-specific risk prediction for recurrence in patients undergoing lobectomy or Sublobar Resection for Small (≤2 cm) Lung Adenocarcinoma: An International Cohort Analysis.
        J thoracic oncol: official pub Intl Assoc Study Lung Cancer. 2019; 14: 72-86https://doi.org/10.1016/j.jtho.2018.09.008
        • Moreira AL
        • Ocampo PSS
        • Xia Y
        • et al.
        A rading system for invasive Pulmonary Adenocarcinoma: A poposal fom the Int Assocfor the Study of Lung Cancer Pathology Committee.
        J thoracic oncol: official publ Intl Assoc Study Lung Cancer. 2020; 15: 1599-1610https://doi.org/10.1016/j.jtho.2020.06.001
        • Nicholson AG
        • Tsao MS
        • Beasley MB
        • et al.
        The 2021 WHO classification of Lung Tumors: impact of Advances since 2015.
        JThoracic Oncol. 2022; 17: 362-387https://doi.org/10.1016/j.jtho.2021.11.003
        • Sereno M
        • He Z
        • Smith CR
        • et al.
        Inclusion of multiple high-risk histopathological criteria improves the prediction of adjuvant chemotherapy efficacy in lung adenocarcinoma.
        Histopathology. 2021; 78: 838-848https://doi.org/10.1111/his.14301
        • Hung J
        • Yeh Y
        • Jeng W
        • et al.
        Predictive value of the international association for the study of lung cancer/American Thoracic Society/European Respiratory Society classification of lung adenocarcinoma in tumor recurrence and patient survival.
        J clinic oncol: official j Am Soc ClinicOncol. 2014; 32: 2357-2364https://doi.org/10.1200/JCO.2013.50.1049
        • Deng C
        • Zheng Q
        • Zhang Y
        • et al.
        Validation of the Novel International Association for the Study of Lung Cancer Grading System for Invasive Pulmonary Adenocarcinoma and Association With Common Driver Mutations.
        J thoracic oncol: official pub Intl AssocStudy Lung Cancer. 2021; 16: 1684-1693https://doi.org/10.1016/j.jtho.2021.07.006
        • Travis WD
        • Brambilla E
        • Noguchi M
        • et al.
        International association for the study of lung cancer/american thoracic society/european respiratory society international multidisciplinary classification of lung adenocarcinoma.
        J thoracic oncol: official pubIntl Assoc Study Lung Cancer. 2011; 6: 244-285https://doi.org/10.1097/JTO.0b013e318206a221
        • Miao Y
        • Zhang J
        • Zou J
        • et al.
        Correlation in histological subtypes with high resolution computed tomography signatures of early stage lung adenocarcinoma.
        Translatl lung cancer res. 2017; 6: 14-22https://doi.org/10.21037/tlcr.2017.02.06
        • Xv Y
        • Lv F
        • Guo H
        • et al.
        A CT-Based Radiomics Nomogram integrated with Clinic-Radiological Features for Preoperatively Predicting WHO/ISUP Grade of Clear Cell Renal Cell Carcinoma.
        Fronti Oncol. 2021; 11https://doi.org/10.3389/fonc.2021.712554
        • Zhang G
        • Xu L
        • Zhao L
        • et al.
        (2020) CT-based radiomics to predict the pathological grade of bladder cancer.
        Europ Radiol. 2020; 30: 6749-6756https://doi.org/10.1007/s00330-020-06893-8
        • Chen X
        • Zhang Y
        • Chen Y
        • et al.
        MRI-based grading of Clear Cell Renal Cell Carcinoma using a Machine Learning Classifier.
        Front Oncol. 2021; 11https://doi.org/10.3389/fonc.2021.708655
        • Peng L
        • Zhang X
        • Liu J
        • et al.
        MRI-radiomics-clinical-based nomogram for prenatal prediction of the placenta accreta spectrum disorders.
        Europ radiol. 2022; 32: 7532-7543https://doi.org/10.1007/s00330-022-08821-4
        • Hou L
        • Wang T
        • Chen D
        • et al.
        Prognostic and predictive value of the newly proposed grading system of invasive pulmonary adenocarcinoma in Chinese patients: a retrospective multicohort study. Modern pathology: an official journal of the United States and Canadian Academy of Pathology.
        Inc. 2022; 35: 749-756https://doi.org/10.1038/s41379-021-00994-5
        • Forest F
        • Laville D
        • Da Cruz V
        • et al.
        WHO grading system for invasive pulmonary lung adenocarcinoma reveals distinct molecular signature: An analysis from the cancer genome atlas database.
        Expmol pathol. 2022; 125104756https://doi.org/10.1016/j.yexmp.2022.104756
        • Landis JR
        • Koch GG
        The measurement of observer agreement for categorical data.
        Biometrics. 1977; 33: 159-174https://doi.org/10.2307/2529310
        • Emoto K
        • Eguchi T
        • Tan KS
        • et al.
        Expansion of the Concept of Micropapillary Adenocarcinoma to include a Newly Recognized Filigree pattern as well as the Classical Pattern Based on 1468 Stage I Lung Adenocarcinomas.
        Journal of Thoracic Oncology. 2019; 14: 1948-1961https://doi.org/10.1016/j.jtho.2019.07.008
        • Moreira AL
        • Ocampo PSS
        • Xia Y
        • et al.
        A gading system for Invasive Pulmonary Adenocarcinoma: aproposal from the International Association for the Study of Lung Cancer Pathology Committee.
        J Thoracic Oncol. 2020; 15: 1599-1610https://doi.org/10.1016/j.jtho.2020.06.001
        • Rokutan-Kurata M
        • Yoshizawa A
        • Ueno K
        • et al.
        Validation Study of the International Association for the Study of Lung Cancer Histologic Grading System of Invasive Lung Adenocarcinoma.
        J thoracic oncol: official pubIntl Assoc Study of Lung Cancer. 2021; 16: 1753-1758https://doi.org/10.1016/j.jtho.2021.04.008
        • Fujikawa R
        • Muraoka Y
        • Kashima J
        Clinicopathologic and Genotypic Features of Lung Adenocarcinoma Characterized by the International Association for the Study of Lung Cancer Grading System.
        Jthoracic oncol: official puIntl Assoc Study Lung Cancer. 2022; 17: 700-707https://doi.org/10.1016/j.jtho.2022.02.005
        • Song SH
        • Park H
        • Lee G
        • et al.
        Imaging Phenotyping using Radiomics to Predict Micropapillary Pattern within Lung Adenocarcinoma.
        J Thoracic Oncol. 2017; 12: 624-632https://doi.org/10.1016/j.jtho.2016.11.2230
        • Park S
        • Lee SM
        • Noh HN
        • et al.
        (2020) Differentiation of predominant subtypes of lung adenocarcinoma using a quantitative radiomics approach on CT.
        EuropRadiol. 2020; 30: 4883-4892https://doi.org/10.1007/s00330-020-06805-w
        • Chen L
        • Yang S
        • Wang H
        • et al.
        Prediction of micropapillary and solid pattern in lung adenocarcinoma using radiomic values extracted from near-pure histopathological subtypes.
        Europ Radiol. 2021; 31: 5127-5138https://doi.org/10.1007/s00330-020-07570-6
        • Bae JM
        • Jeong JY
        • Lee HY
        • et al.
        Pathologic stratification of operable lung adenocarcinoma using radiomics features extracted from dual energy CT images.
        Oncotarget. 2017; 8: 523-535https://doi.org/10.18632/oncotarget.13476
        • Xu Y
        • Ji W
        • Hou L
        • et al.
        Enhanced CT-Based Radiomics to predict Micropapillary pattern within Lung Invasive Adenocarcinoma.
        Front oncol. 2021; 11704994https://doi.org/10.3389/fonc.2021.704994
        • Lubner MG
        • Stabo N
        • Abel EJ
        • et al.
        CT Textural analysis of Large Primary Renal Cell Carcinomas: pretreatment Tumor Heterogeneity correlates with Histologic Findings and Clinical outcomes.
        AJR Am J Roentgenol. 2016; 207: 96-105https://doi.org/10.2214/AJR.15.15451
        • Skogen K
        • Ganeshan B
        • Good C
        • et al.
        Measurements of heterogeneity in gliomas on computed tomography relationship to tumour grade.
        Journal of Neuro-Oncology. 2013; 111: 213-219https://doi.org/10.1007/s11060-012-1010-5
        • Ganeshan B
        • Goh V
        • Mandeville HC
        • et al.
        Non-small cell lung cancer: histopathologic correlates for texture parameters at CT.
        Radiology. 2013; 266: 326-336https://doi.org/10.1148/radiol.12112428
        • Lu C
        • Shiradkar R
        • Liu Z
        Integrating pathomics with radiomics and genomics for cancer prognosis: abrief review.
        Chinese jcancer res = Chung-kuo yen cheng yen chiu. 2021; 33: 563-573https://doi.org/10.21147/j.issn.1000-9604.2021.05.03
        • Naidich DP
        • Bankier AA
        • MacMahon H
        • et al.
        Recommendations for the management of subsolid pulmonary nodules detected at CT: a statement from the Fleischner Society.
        Radiology. 2013; 266: 304-317
        • Hansell DM
        • Bankier AA
        • MacMahon H
        • et al.
        Fleischner Society: glossary of terms for thoracic imaging.
        Radiology. 2008; 246: 697-722
      2. Li F, Sone S, Abe H, et al. Malignant versus benign nodules at CT screening for lung cancer: comparison of thin-section CT findings. 2004 793–798..

        • Lee E
        • Son D
        • Kim S
        • et al.
        Prediction of recurrence-free survival in postoperative non-small cell lung cancer patients by using an integrated model of clinical information and gene expression.
        Clinic cancer research: official j AmAssoc Cancer Res. 2008; 14: 7397-7404