Introduction
Although lobectomy is the standard treatment for early-stage, non–small-cell lung cancer, limited resection is emerging as a surgical alternative and has proven to be non-inferior to lobectomy in terms of overall survival while preserving lung function [1-3]. However, the 5-year recurrence rate after a limited resection is reported to be twofold compared with lobectomy. This adverse outcome of early-stage lung cancer can be attributed to intratumoral heterogeneity and differences in prognostic outcome according to histologic and molecular subtypes in both adenocarcinoma and squamous cell carcinoma [4-8]. The histologic and molecular characterization of squamous cell carcinoma remains elusive for both prognostication and therapeutic targeting [9,10]. In contrast, copious literature describes prognostic outcomes for each histologic pattern of lung adenocarcinoma, including a reported 10-year recurrence-free survival rate of 100% in minimally invasive adenocarcinoma and adenocarcinoma in situ after surgery [11,12]. Current TNM staging and its supporting material offer well-documented information regarding patient prognosis, however, this staging system does not provide enough information specifically about individual patient prognosis depending on emerging treatment options. Given that background, there is growing attention to the prognostic significance of high-grade lung adenocarcinoma patterns, especially the micropapillary and solid (MPsol) pattern [9,13,14]. Poorly differentiated histology is classified as the highest grade in the 2021 World Health Organization classification [9], and though it is not included in the tumor grading system, spread-through-air-space (STAS) also correlates with high recurrence and low survival rates [15-17]. Patients with those high-grade histologic patterns are known to show different responses to treatment than those without them [18] and to benefit from aggressive surgical resection [19,20]. A high proportion of high-grade patterns within a tumor correlates with a high tumor mutational burden. In addition, the presence of STAS is reported to be associated with high-grade patterns in each tumor (especially the micropapillary pattern) and intrathoracic recurrence [21].
Because accurate preoperative prediction of specific histologic types could help establish personalized treatment plans in even early-stage lung adenocarcinoma, this study focuses on preoperative quantitative imaging identification of high-risk pathologic features (MPsol, STAS, and poor differentiation) in lung adenocarcinoma. Attempts to elucidate quantitative imaging characteristics and link them to high-risk pathologic features have been made in several studies using various imaging modalities. According to a study by Lee et al. [22], an irregular tumor margin was associated with poor overall survival times and the micropapillary pattern. Other studies reported correlations between SUVmax (maximum standardized uptake value) in 18F-fluorodeoxyglucose (FDG) positron emission tomography (PET)/computed tomography (CT) [23] and magnetic resonance imaging (MRI)–based radiomics features [24] that could be used to predict histologic high-grade lung adenocarcinoma. Song et al. [25] explored radiomics features related to the MPsol pattern. For STAS pattern, irregular and lobulated tumor margin was reported to be associated with presence of STAS [17]. CT-based radiomics approaches have also been used for imaging prediction of STAS [26,27]. Therefore, in this study, we take a radiomics approach to explore quantitative MRI features that might reflect high-risk pathologic features for lung adenocarcinoma. Also, as prospective study, we attempt to screen high-risk lung cancer patients with our proposed imaging criteria.
Materials and Methods
1. Patients and screening step for high-risk lung cancer patients
Between August 2018 and January 2020, we identified patients diagnosed with clinical N0 lung cancer who were candidates for curative surgical resection at Samsung Medical Center, Korea. Here, clinical N0 was evaluated by images (CT and 18F-FDG PET/CT). Among them, patients who underwent both contrast-enhanced chest CT and 18F-FDG PET/CT were selected. We designed an imaging criterion to screen lung cancer patients who are more likely to have high-risk pathologic features and therefore benefit from additional lung MRI. Our previous studies reported that an irregular tumor margin and high uptake in 18F-FDG PET/CT correlated positively with a poor prognosis and high-grade histology, respectively [22,23,28]. A spiculated margin was reported to correlate with STAS [17]. In addition, purely solid-appearing lung adenocarcinoma was reported to show poorer prognosis than part-solid nodules [29], a high association with the MPsol subtype [5,30], and a high prevalence of nodal metastasis [31]. With this background, we set purely solid lesions with either (1) a suspicious irregular or poorly defined margin in CT with a SUVmax greater than 5 or (2) a definite irregular or poorly defined border with a SUVmax between 3 and 5 as the high-risk features (Fig. 1). In this study, the determination of a ‘suspicious’ or ‘definite’ irregular or poorly defined margin was based on whether 25% or more of the tumor surface exhibited irregularity or poorly defined margin upon three-dimensional volume rendering. Specifically, tumors with 25% or more of the surface classified as irregular or poorly defined margin were categorized as ‘definite,’ while those with less than 25% were classified as ‘suspicious.’ The patients who met those imaging criteria were asked to take preoperative lung MRI, and were grouped as magnetic resonance (MR)–eligible group. Patients who did not meet the above-mentioned criteria were grouped as MR-non-eligible group and did not undergo lung MRI.
2. Image acquisition: CT, 18F-FDG PET/CT, and MRI
Contrast-enhanced chest CT scans were obtained for all patients, with 80 mL of contrast material at 2 mL/sec followed by 20 mL of normal saline at 2 mL/sec. Various CT scanners manufactured by different vendors were used and some patients took CT scans in outside hospitals. Details of image acquisition are described in our previous study [32].
18F-FDG PET/CT was scanned using Discovery STe (GE Healthcare, Milwaukee, WI) or Discovery MI-DR (GE Healthcare) scanner. CT scanning was performed with a continuous spiral technique and followed by PET scanning from base of skull to mid-thigh. Images from outside hospitals were reviewed again in our hospital.
All MRI examinations were performed with a 3 Tesla (T) imager (Skyra, Siemens, Erlangen, Germany) using surface array coils with 32 receiving channels. The imaging protocol consisted of axial and coronal STIR (short tau inversion recovery) images [33], a T1 volumetric interpolated breath-hold examination with contrast (T1 VIBE contrast), and diffusion-weighted image (DWI). Detailed imaging parameters are described in Supplementary Materials.
After MRI image acquisition, regions of interest (ROIs) were drawn for the tumor in the STIR and T1 VIBE contrast-enhanced axial images. A semi-automatic approach in AVIEW software (Coreline Soft, Seoul, Korea) was used for drawing the ROIs. The ROIs were then modified to fit the tumor by one thoracic radiologist (Jh.K. with 5 years of experience) and ROIs were reviewed by one senior thoracic radiologist (H.Y.L. with 17 years of experience). Apparent diffusion coefficient (ADC) value was measured by placing the ROI over the entire tumor. All ROIs were double-checked by the corresponding author.
3. Pathologic analysis
Tumor size, histologic grade staging, and vascular/lymphatic/pleural/extra-pleural invasion were reported for each specimen. Histologic subtyping was reported following 2011 lung adenocarcinoma classification of the International Association for the Study of Lung Cancer (IASLC)/American Thoracic Society (ATS)/European Respiratory Society (ERS) and slides were reviewed again to ensure compliance with newly proposed grading system by the 2021 IASLC pathology committee [9]. Further elaboration for pathology analysis is available in Supplementary Materials.
4. Radiomics feature extraction
54 radiomics features were automatically extracted from ROIs of T1- and T2-weighted images (-WI) using the PyRadiomics open-source software package implemented in Python (https://pyradiomics.readthedocs.io/en/latest/) [34]. Detailed explanation for extracted radiomics features are available in Supplementary Materials and S1 Table.
5. Statistical analysis
The comparison of baseline characteristics between the MR-eligible and non-eligible groups was performed using T-testing for continuous variables and chi-square testing for categorical variables. Kaplan-Meier plot was drawn to compare 4-year disease-free survival and 4-year overall survival of the two groups.
We calculated the variance inflation factor (VIF) and any features showing high correlation (VIF > 10) were trimmed out. Univariable analyses (T tests or chi-square tests) were run for the clinical and radiomics variables. After trimming out the highly correlated features, we took the radiomics variables with p-values less than 0.2 and applied the least absolute shrinkage and selection operator (LASSO) for classification modeling. Regardless of their p-values, age, sex, and smoking status were always included in the clinical variables for model building.
Using the features selected from LASSO, we built prediction models and tuned them using 5-fold cross-validation in the training set. Among the resulting five prediction models, we chose the model with the best performance and evaluated its performance in the test sets. Accuracy, area under the receiver operating curve (AUC), sensitivity, specificity, and mean AUC were calculated for all cross-validations. All statistical analyses for modeling were performed using the Statistics and Machine Learning Toolbox in Matlab and in-house code (The MathWorks, Natick, MA). General statistical analysis was conducted using R software ver. 4.3.2 (R Core Team 2021, R Software, R Foundation for Statistical Computing, Vienna, Austria), and SPSS ver. 22.0 (IBM Corp., Armonk, NY).
Results
1. Demographics of the study cohort
We recruited 616 clinical N0 patients, of whom 174 met the criteria for high-risk features in their CT and 18F-FDG PET/CT images. Excluding patients who could not or did not want to undergo MRI, 72 patients underwent lung MRI before curative resection. All 616 patients underwent curative resection, with 467 (75.8%) receiving a lobectomy with or without a combined wedge resection and 148 (24.0%) undergoing only a wedge resection. After surgery, 524 cases (85.1%) were confirmed as lung adenocarcinoma, including 55 patients who had undergone lung MRI (Fig. 1).
The demographics of the MR-eligible and MR-non-eligible groups are summarized in Table 1. The MR-eligible group showed higher clinical T staging, with a statistical significantly larger size on CT and higher SUVmax than the MR-non-eligible group (p < 0.001). Only clinical stage I or II patients were included in this study, and pathologic stages IA1, IA2, and IA3 accounted for 11.6% (n=61), 39.1% (n=205), and 25% (n=131), respectively, were confirmed as lung adenocarcinoma. In addition, 55 (10.5%) cases were diagnosed as p1B, 33 (6.3%) as pII, 36 (5.9%) as pIII, and 3 as pIV. Upstaging was more common in the MR-eligible group than the MR-non-eligible group (40.6% vs. 26.9%, p=0.003). Nodal upstaging was confirmed in 56 cases of lung adenocarcinoma, and that proportion was higher in the MR-eligible group (29.2% vs. 4.2%, p < 0.001). Vascular (6.4% vs. 2.0%, p=0.007) and lymphatic invasion (21.4% vs. 6.1%, p < 0.001) were confirmed in a higher proportion of the MR-eligible group than the MR-non-eligible group.
The MR-eligible group showed higher proportions of the STAS (25.3% vs. 11.3%), micropapillary pattern (50.0% vs. 15.8) and solid pattern (29.7 vs. 7.0) than the MR-non-eligible group (p < 0.001). Moreover, MR-eligible-group showed lower 4-year disease-free survival (60.9% vs. 76.6%, p < 0.001) (Fig. 2) and 4-year overall survival (69.5% vs. 78.5%, p=0.02) (Fig. 3) compared with MR-non-eligible group. Among the 55 MR-eligible cases of surgically proven lung adenocarcinoma, the MPsol, poorly differentiated, and STAS pathologic patterns were found in 38, 13, and 19 cases, respectively.
2. Pathology-radiology correlation
As reported in the previous literature, high-risk pathologic features correlated with strong FDG uptake in 18F-FDG PET/CT images and low ADC values upon pictomicrograph comparison. Fig. 4 shows a lung adenocarcinoma pictomicrograph and images. The tumor portion containing the micropapillary pattern and STAS shows increased FDG uptake (SUVmax 3.5) and a low ADC value. The tumor shown in Fig. 5 has homogeneously strong FDG uptake (SUVmax 17.3) in the 18F-FDG PET/CT image, but only the portion with a solid component and STAS shows a low ADC value in the MRI.
3. Prediction of high-risk pathologic features using MR-radiomics features
Radiomics features from both T1- and T2-WI were selected for all high-risk pathologic features. The selected radiomics features are presented in Table 2. MPsol correlated positively with flatness and zone entropy, and correlated negatively with cluster shade and difference variance. STAS correlated positively with the Large Area Low Gray Level Emphasis (LALGLE) (T2WI) feature. Poor differentiation (2011 grading system) was positively associated with correlation (T1 weighted image, T1WI) and Imc2 (T2WI). It was negatively associated with cluster shade on T2WI. The 2021 poorly differentiated pattern grading system showed association with different radiomics features. It correlated positively with 10&90 percentile and range and all first-order features in the T1 contrast-enhanced images. It correlated negatively with 75th percentile of slope (T1WI), cluster shade (T2 weighted image, T2WI), and low gray level zone emphasis (T2WI).
The prediction power for MPsol, STAS, and the poorly differentiated pattern defined by both the 2011 and 2021 grading systems was good, with mean AUC values of 0.7512, 0.7631, 0.7976, and 0.8857, respectively in the test set (Table 3). We added the clinical variables including age, sex, smoking status, and other variables with p-values of less than 0.2 in the univariate analyses, and the model performance for MPsol and the poorly differentiated pattern of 2021 grading system improved (0.7512 to 0.8595 and 0.8857 to 0.9071, respectively). Included clinical variables were age, sex, and smoking and these clinical variables were always included for model building regardless of univariate analysis result.
Discussion
The key findings of our study can be summarized as (1) a multimodal approach to building a prediction model for MPsol, (2) a novel two-step approach in a prospective setting to set imaging criteria for high-risk histologic patterns that can effectively screen patients who might benefit from an additional lung MRI, and (3) an effective MR-based radiomics prediction model for STAS.
The MPsol component in lung adenocarcinoma correlates with a high rate of tumor recurrence, poor overall survival time, and nodal upstaging [5,28,35,36]. Although limited resection in early-stage lung cancer with MPsol is reported to be non-inferior to lobectomy, complete lymph node dissection and adjuvant chemotherapy are associated with better outcomes than limited resection alone [35,37,38]. Many previous studies have used radiomics or deep-learning approaches with CT images to predict MPsol components, with reported AUC values between 0.73 and 0.82 in the validation sets [6,25,39-42]. To our knowledge, one previous MRI-based report predicted histologic grades using a radiomics approach with T2WI and DWI together [24], and in our previous study, we tested a multimodal clustering approach to determine imaging-based intratumoral heterogeneity and tumor aggressiveness [43]. However, no previous multimodal approach was specifically targeted to predict MPsol patterns. Our prediction model using a multimodal approach shows high predictive power, with an AUC value of 0.8929, which is superior to previous study results.
Upon analysis with the relevant radiomics features, MPsol correlated positively with zone entropy and negatively with cluster shade and difference variance. Cluster shade measures the skewness of the co-occurrence matrix, and a higher value implies less clustering of voxels with similar values. The negative correlation we found could reflect both subclonal expansion and diversity: during tumor evolution, the micropapillary pattern exhibits a strong correlation with subclonal diversity, whereas the solid pattern demonstrates a strong correlation with subclonal expansion [21]. These factors together can result in both tumoral heterogeneity and clustering in our study result. The negative correlation with Difference Variance can be interpreted in the same context. A small MPsol component within the tumor can affect survival and recurrence, so preoperative imaging prediction of MPsol components can change treatment plans of a patient. Our study shows that our multimodal approach and machine-learning-based prediction model offers excellent diagnostic predictability, demonstrating the clinical utility in real-world practice.
Lung MRI is not a conventional diagnostic imaging modality for lung cancer because it is more expensive and takes longer than 18F-FDG PET/CT and CT; therefore, it is crucial to effectively screen patients who will benefit most from additional imaging study. In this prospective study, patients meeting our imaging criteria (MR-eligible group) showed a higher incidence of nodal metastases, lymphatic invasion, and high-risk pathologic features, than the patients who did not meet our criteria (MR-non-eligible group) (Table 1). This demonstrates the effectiveness of our imaging criteria for screening high-risk patients. Currently, no guidelines are available for screening high-risk patients preoperatively and recommending lung MRI. Therefore, these imaging criteria will help clinicians select patients who will benefit from lung MRI for tissue characterization.
Lung adenocarcinomas have heterogenous histology, and the conventional grading system does not take small portions of poor prognostic patterns within tumors into account. In contrast, the new IASLC grading system proposed in 2021 reflects tumor heterogeneity because even a small portion of a poor prognostic histology pattern such as MPsol can have a negative effect on a patient’s overall prognosis. In this study, all slides were reviewed based on both the old and new grading systems, and 76 more patients were classified as having the poorly differentiated pattern by the new grading system. No down-grading from the old system to the new one occurred, and the percentage of patients up-graded was 14.5% (76 out of 524). The portion of up-graded patients was higher in the MR-eligible group than the MR-non-eligible group (38.2% vs. 11.7%). We built our MRI-based prediction model based on each grading system, and the predictive power was better when using the new grading system (mean AUC 0.7601 vs. 0.9071, respectively, in the 5-fold training sets). By taking advantage of the better prediction of clinical outcomes offered by the new grading system, our prediction model can support both clinicians and radiologists in characterizing tumors.
The presence of STAS is significantly related to local tumor recurrence and the presence of MPsol components [44]. Although not included in the final model of the new IASLC pathologic grading approach, STAS alone showed good correlation with poor prognosis, with an AUC value of 0.752 for recurrence and 0.765 for survival [13]. In addition, compared with lobectomy, limited resection was associated with recurrence and lung cancer–specific death in patients with STAS [19,45,46]. STAS is also related to occult lymph node metastasis in clinical stage IA patients [47]. In this regard, previous research used radiomics for preoperative prediction of a STAS component and reported AUC values of 0.69 to 0.85 [27,48-50]. In our study, the AUC value of our best prediction model in the 5-fold test sets was good at 0.7631. In our analysis with MRI-based radiomics features, STAS correlated positively with LALGLE on T2WI, which represents the solid portions of tumors with high cellularity. During tumor growth, both highly cellular and necrotic areas expand at the same time. According to CT-based radiomics studies, both LALGLE and Large Area High Gray Level Emphasis (LAHGLE) are known to increase in value. Though CT and MR are different imaging modalities, tumor characteristics reflect radiomics features and in our study, LALGLE can be interpreted as solid cellular portions of tumor. Whereas previous research was all based on CT images, our novel multimodal-based radiomics approach showed comparable predictive power for STAS. Our approach can help clinicians make optimal surgical plans, such as avoiding limited resection.
Our study has several limitations. First, we included a relatively small number of patients for analysis and did not conduct external validation. Lung MRI is not a conventional diagnostic modality for lung cancer, and it is difficult to recruit patients for research purposes. Even with this limitation, our study serves as a cornerstone for future MR-radiomics studies, presenting MRI as potential preoperative diagnostic tool that can predict important pathologic patterns and we expect future studies with larger patient pool and external validation. Second, as our screening process solely relies on imaging, there is a risk of including non-adenocarcinoma lung cancer patients in MR-eligible group, leading to unnecessary MR imaging. We anticipate that future or ongoing artificial intelligence studies accurately predicting lung cancer histology to complement our two-step process, creating a synergistic effect and enhancing the efficiency of patient selection for MR imaging. Lastly, although we had many available sequences for MRI, only T1 contrast-enhanced, T2 STIR, and ADC map images were used for analysis. Also, all the recruited patients had available CT and PET data; however, no data other than SUVmax were used in this study. The incorporation of available CT and PET data and the inclusion of more MR sequences could enhance study outcomes. Therefore, we are planning a multimodal approach to high-risk histologic marker prediction for our follow-up study.
In this prospective study to incorporate MRI in predicting high-risk pathologic features in lung adenocarcinoma, our imaging criteria successfully selected patients with a high chance of having high-grade histologic patterns. In addition, our prediction model for high-risk pathologic features using MR-based radiomics variables in T1- and T2-WI had excellent predictive power. Our proposed imaging criteria will help clinicians select optimal candidates for lung MRI, and our MR-based prediction model will enable detailed preoperative evaluations of histopathologic status, particularly indicating high-risk patients. These screening criteria and our prediction model can help clinicians decide precise treatment plans for patients with lung cancer.
PDF Links
PubReader
ePub Link
XML Download
Full text via DOI
Download Citation 
Supplement1
Print 
