Evaluating Solid Lung Adenocarcinoma Anaplastic Lymphoma Kinase Gene Rearrangement Using Noninvasive Radiomics Biomarkers

Introduction

Lung cancer is one of the leading causes of cancer-related death worldwide, among which approximately 85–90%¹ of cases are non-small cell lung cancer (NSCLC). Lung adenocarcinomas are marked with distinct genomic changes that are correlated with clinically relevant molecular subtypes.² Nowadays, the treatment of advanced NSCLC including adenocarcinoma has shifted into determining oncogenic driver mutation subtypes.³ Epidermal growth factor receptor (EGFR) mutations and anaplastic lymphoma kinase (ALK) rearrangement are the two most targeting “druggable” oncogenic driver mutation subtypes in NSCLC.⁴ ALK+ NSCLC, though it accounts for only 2–7% of NSCLC,^5,6 expresses a sensitive responsiveness to ALK inhibitors such as ceritinib, crizotinib, lorlatinib,⁷ and brigatinib.⁸ ALK targeted therapies have been developed and subsequently continuously evolved for more than ten years since the discovery of ALK rearrangement in NSCLC, boosting achievement in median progression free survival (PFS),⁹ overall response rate (ORR),¹⁰ overall survival (OS),¹¹ and individual precision medicine.⁷

To determine eligible pretreated adenocarcinoma patients for ALK targeted therapy, it is crucial to address both pragmatic and cost-efficient assessment of ALK status. Fluorescence in situ hybridization (FISH) and next-generation sequencing (NGS) are recognized as the “gold standard” for ALK gene test.¹² Immunohistochemistry (IHC) serves as a widely-used surrogate to FISH or molecular method. Other methods include quantitative reverse transcription-PCR, RNA scope, etc. Traditional invasive gene testing was recognized as the one and only effective method to identify those ALK positive patients. The incapability to obtain overall information about highly heterogeneous lung cancer preoperatively remains a major limitation of traditional gene testing based on invasive local puncture sample. Besides, the broad employment of NGS in routine clinical setting is hampered by its relatively time consuming, invasive nature, as well as complicated processes.¹² All these limitations may enormously restrict the practicability of individual precision medicine.

In clinical practice, computed tomography (CT) scanning is the most easy-to-handle approach to diagnose and evaluate pretreated lung cancer. Radiomics is an emerging field that involves the use of high-throughput computing to extract a large number of quantitative features from medical imaging data.¹³ Recently, radiomics research interest in NSCLC including adenocarcinoma has been growing, involving survival prediction,^14,15 distant metastasis prediction,¹⁶ local control and radiation pneumonitis,¹⁷ histologic subtypes classification,¹⁸ targeted therapy response phenotype,¹⁹ etc. The attempts to define the association between imaging textures (CT, for example) and cancer genomic phenotypes have accelerated the development of “radiogenomics”. By extracting high throughout of CT series, radiogenomics allows the noninvasive diagnosis of tumor gene mutations in NSCLC EGFR mutation prediction,²⁰ prediction of pathological stage in NSCLC using machine learning²¹ or deep learning²² and NSCLC multi-subtype classifications.²³

Whilst several previous studies have demonstrated certain radiological characteristics associated with NSCLC ALK+ and built a diagnostic model²⁴ based on radiological characteristics and clinical factors, the role of CT based radiogenomics machine learning in ALK+ solid lung adenocarcinoma remains to be explored. Therefore, we hypothesized that lung adenocarcinoma ALK status differences and relevant phenotype discriminations are mirrored in radiomics features. Thus, radiogenomics machine learning of standard enhanced CT scans, which combines 3D radiomics features, blood supply distinction, and entire tumor burden information, can help to identify solid lung adenocarcinoma anaplastic lymphoma kinase gene rearrangement status.

Materials and Methods

Figure 1 demonstrates the flowchart of this study.

Figure 1 Flowchart of this study. (A) Patients. (B) Image Acquisition. (C) ROI Segmentation. (D) Feature Extraction and Selection. (E) SVM Construction and Validation.

Ethical Approval

This study was conducted in accordance with the Declaration of Helsinki. This retrospective single center analysis was approved by the local ethics committee of Zhejiang Cancer Hospital (IRB-2019-142). Written informed consent for pre- and post-contrast enhanced CT series were obtained from all patients.

Patients

We retrospectively reviewed all pretreated solid lung adenocarcinoma patients whose ALK and EGFR status were analyzed and who underwent pre- and post-contrast CT from September 2015 to September 2018. Accordingly, forty-seven ALK+ patients were identified as the study group. Meanwhile, pretreated solid adenocarcinoma patients with positive EGFR mutations (n=44) and none mutation (n=49) were randomly selected from local database on a 1:1:1 ratio as contrast group. In total, 140 cases were initially gathered (53 men and 87 women; mean age, 54.19 years; range, 23–80 years) according to the following criteria: 1) pathologically confirmed adenocarcinoma; 2) presence of solid lesion on pretreatment CT; 3) ALK rearrangement was diagnosed by IHC or NGS technologies; EGFR mutation was tested by amplification refractory mutation system (ARMs). 4) Less than one-month interval between CT exams and gene alternation analysis. 5) No compound cancer history. In classifier two, cases with inconsistency between pre- and post-contrast slice thickness were further excluded. Clinical characteristics including sex, age, histopathology, staging, and smoking history were recorded. Clinical features between ALK+ group and ALK- group were analyzed using SPSS 18.0 for windows (SPSS Inc); the Pearson χ2 test or Fisher’s exact test was used to compare qualitative variables; and quantitative variables were analyzed by the t-test or Pearson’s correlation test.

Image Acquisition

All patients underwent both pre- and standard iodine enhanced CT scans when they were initially admitted to our institution before any treatment. CT exams were performed either on a 16-detector CT scanner or on a 64-detector CT scanner (Sensation 16; SOMATOM Definition Flash 64; Siemens Medical Solutions, Forchheim, Germany). For the 16-detector scanner, the parameters were 120 kV and 100 effective mA. For the 64-detector scanner, the parameters were 120 kV and 110 effective mA. Dose modulation was used in both scanners. Images were reconstructed with a section thickness of 5 mm and a gap of 1 mm. Post-contrast scanning started 38–48 s after intravenous administration 100 mL of loversol injection (Optiray, Guerbet, Villepinte, France) at a rate of 2.8 mL/sec by using a power injector.

Image Pre-Processing and Segmentation

Every lesion was segmented manually slice by slice by two radiologists (reader 1: X.Y.G. with 5 years experience; reader 2: A.N.Z. with 5 years experience) using ITK-SNAP (http://www.itksnap.org), and then scrutinized by a radiology professor (H.Z.Z, with 26 years work experience). We further tested intra- and inter-observer reproducibility according to Wu et al.²⁵ Accordingly, we initially chose 25 lesions and lay an ICC greater than 0.75 as satisfying reproducibility agreement.

Machine Learning Process

In order to be potentially clinically effective, we designed classifiers on pre-contrast CT features (classifier 1) and standard post-contrast enhanced CT features (classifier 2).

In each classifier, radiomics features were extracted from each ROI with Pyradiomics (http://pyradiomics.readthedocs.io/en/latest/index.html). Classes of features included Shape, First Order, Gray Level Co-occurrence Matrix, Gray Level Size Zone Matrix, Gray Level Run Length Matrix, Neighboring Gray Tone Difference Matrix and Gray Level Dependence Matrix. We then applied normalization and standardization on the feature matrix. For each feature vector, we calculated the L2 norm and divided by it. Then the feature vector was mapped to a unit vector. Since the dimension of feature space was high, we compared the similarity of each feature pair. According to a previous study,²⁶ if the Pearson correlation coefficient (PCC) value of the feature pair was larger than 0.86, we randomly removed one of them. After this process, the dimension of the feature space was reduced and each feature was independent to each other. Before building the model, we tried analysis of variance (ANOVA) and recursive feature elimination (RFE) separately to select features. Finally, we picked the best-behaved feature selection method by comparing the results between the two methods. ANOVA is a common method to explore the significant features corresponding to the labels. F-value was calculated to evaluate the relationship between features and the labels. We sorted features according to the corresponding F-value and selected a specific number of features to build the model. The goal of RFE is to select features based on a classifier by recursively considering smaller set of the features. We used support vector machine (SVM) as the machine learning classifier and applied synthetic minority oversampling technique (SMOTE) algorithm to balance between data groups. The kernel function has the ability to map the features into a higher dimension to search the hyper-plane for separating the cases with different labels. Here we used the linear kernel function because it was easier to explain the coefficients of the features for the final model. To prove the performance of the model, we applied cross-validation with 5-folder on the data set. The performance of the model was evaluated using receiver operating characteristic (ROC) curve analysis. The area under the ROC curve (AUC) was calculated for quantification. The accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were also calculated at a cutoff value that maximized the value of the Yorden index. We also boosted estimation 1000 times and applied paired t-test to give the 95% confidence interval. All these processes were implemented with FeAture Explorer (FAE, v0.2.5, https://github.com/salan668/FAE) on Python (3.6.8, https://www.python.org/).

Results

Clinical and Imaging Data

Table 1 demonstrates the clinical characteristics of ALK+ adenocarcinoma group and ALK- adenocarcinoma group. In classifier two, 3 ALK+ patients and 13 ALK- patients (6 EGFR+, 7 negative) were excluded due to inconsistency between pre- and post-contrast slice thickness. All the inter- and intra-observer reproducibility of lesions’ segmentation reached a satisfying standard. Hence, the rest of the ROIs were drawn by the first radiologist. Figure 2 demonstrates an example of ALK+ patients with pre- and standard post-contrast CT scans.

Table 1 Clinical Characteristics of the ALK (+) and ALK (-) Patients

Figure 2 CT images obtained in a 49-year old woman with ALK rearrangement solid lung adenocarcinoma. Transverse mediastinal window, 5 mm slice thickness. (A and B) Pre-contrast scans. (C and D) Standard post-contrast scans.

Classifier 1: Pre-Contrast CT Radiogenomics Classifier

During this process, only features extracted from pre-contrast CT were analyzed. One hundred and forty cases were randomly split into training data set (n=98), and independent testing data set (n=42) on a 7:3 ratio. The top 25 features were chosen by ANOVA across a total of 851 features. Figure 3B demonstrates the OSE feature selection method. In this classifier, the pre-contrast CT radiogenomics classifier provided a diagnostic accuracy of 78.57%. The AUC was 85.90% in the training cohort, 80.10% in the validation cohort, and 80.10% (95% CI: 0.6538–0.9222) in the test cohort with 71.43% sensibility, 82.14% specificity, 66.67% PPV, and 85.19% NPV (Table 2). Figure 3A demonstrates the ROC curve of classifier 1. Figure 3C demonstrates the contribution of selected features in classifier 1.

Table 2 Diagnostic Efficiency of Classifier 1

Figure 3 Results of classifier 1. (A) The receiver-operating characteristic curve of classifier 1. (B) Selection of optimal features using ANOVA. (C) Contribution of selected radiomics features of classifier 1.

Classifier 2: Post-Contrast CT Radiogenomics Classifier

During this process, only features extracted from post-contrast CT were analyzed. One hundred and twenty-four cases were randomly split into training data set (n=87), and independent t-testing data set (n=37) on a 7:3 ratio. The top 19 features were chosen by RFE across a total of 851 features. Figure 4B demonstrates the OSE feature selection method. In this classifier, the pre-contrast CT radiogenomics classifier provided a diagnostic accuracy of 81.08%. The AUC was 91.10% in the training cohort, 82.85% in the validation cohort, and 82.90% (95% CI: 0.6630–0.9567) in the test cohort with 76.92% sensibility, 83.33% specificity, 71.43% PPV, and 86.96% NPV (Table 3). Figure 4A demonstrates the ROC curve of classifier 2. Figure 4C demonstrates the contribution of selected features in classifier 2.

Table 3 Diagnostic Efficiency of Classifier 2

Figure 4 Results of classifier 2. (A) The receiver-operating characteristic curve of classifier 2. (B) Selection of optimal features using RFE. (C) Contribution of selected radiomics features of classifier 2.

Discussion

Highlighted by therapeutic advances in ALK inhibitors, assessing personalized ALK status effectively and accurately is of significant importance. However, in clinical practice, the availability of attainable and sufficient ALK status test is sometimes challenging. In this study, a radiogenomics machine learning classifier built with 19 selected standard post-contrast CT radiomics feature to assess ALK+ status in pretreated solid lung adenocarcinoma showed the best diagnostic performance, compared with that of pre-contrast CT classifier. Therefore, the standard post-contrast CT machine learning radiogenomics classifier could help precisely and non-invasively identify solid lung adenocarcinoma ALK rearrangement status, which may act as a pragmatic and cost-efficient substitute for traditional invasive ALK status test.

Generally, there are three major oncogenic driver mutation subtypes of lung adenocarcinoma, namely the most common and widely assessed epidermal growth factor receptor (EGFR), ALK and Kirsten rat sarcoma viral oncogene homolog (KRAS). ALK+ cases represent a small but significant adenocarcinoma subgroup, because it makes up to 30% in the nonsmoking EGFR- NSCLC population.²⁴ Notably, among all oncogenic driver mutation subtypes, only ALK+ and EGFR+ cases reacted to specific targeted drugs approved by both FDA^27,28 and CFDA strikingly. The latest guideline²⁹ suggested that EGFR and ALK gene status must be tested in lung cancer. ALK fusions typically occur independently from EGFR mutations,³⁰ making it important to distinguish ALK+ cases from EGFR+ cases. Hence, in this study, we enrolled both EGFR+ cases and ALK- EGFR- oncogenic driver mutation cases in the contrast cohort.² It is widely recognized that ALK+ adenocarcinomas are usually solid nodules or masses,² pure ground glass nodules are scarce³¹ and have distinct radiological characteristics. Besides, conventional percutaneous transthoracic needle biopsy for pure ground glass nodules has certain limitations³² and was not highly recommended in our institution. Thus, we focused on solid lung adenocarcinoma lesions in this study.

In contrast to overall cases, ALK+ patients are typically younger, and present a non-smoking (defined as those who smoked less than 100 cigarettes in their lifetime³³) or relatively light smoking history.³⁴ Consistent with previous studies, ALK+ patients were significantly younger compared with ALK- patients in our study. However, smoking history and female gender were not significantly distinguishable. Most patients in our ALK+ study group had distant metastasis. Common metastatic sites in our ALK+ study group were brain, pleura, contralateral lung, liver, and bone.

In clinical practice, IHC serves as a practical method to detect ALK status. Generally, only samples with equivocal IHC patterns were further handled with the “gold standard” FISH or NGS methods. In our research, all the patients underwent IHC to test ALK status, only four patients with not high positive ALK expression were further diagnosed with ALK+ by NGS test. In addition to obtaining traditional tissue, recently, several minimally invasive cytological-based approaches have been adopted³⁵ including samples from fine-needle aspiration,³⁶ pleural effusion, bronchial brushing, etc. However, multiple obstacles such as relatively high cost, invasive risks, limited evaluation of whole lesion area as well as limited sensitivity cytological samples³⁷ hamper the widespread identification of ALK+ adenocarcinoma patients. Radiogenomics addresses the need for a noninvasive measurement to distinguish patients with ALK+ adenocarcinoma who should undergo ALK targeted therapy. In addition, this radiogenomics analysis adequately involved 3D radiomics features, blood supply distinction and entire tumor burden information. Besides, clinical CT series are more accessible since most patients tend to be clinically diagnosed by CT scanning firstly.

Recent ALK+ NSCLC radiology studies mainly focused on subjective traditional radiological characteristics analysis³⁸ based on naked eyes or on quantitative parameters drained from dual-energy spectral computed tomography (DESCT).³⁹ Kim et al²⁴ built an ALK+ predictive model yielding an AUC value of 0.832, combining multi-modality radiological characteristics and clinical factors. Recently, Ruan et al reported a best AUC value of 0.873, combining several ¹⁸F-FDG PET/CT parameters.³⁷ Only a limited, small amount of readable imaging features were evaluated and utilized in recent conventional studies, leaving a huge amount of precious image information which could be potentially helpful to diagnose ALK+ status, untouched. By combining both various radiomics features of CT and novel machine learning, this study mined the conventional but practical CT image data to the utmost extent. This preliminary study outperformed previous studies by reaching comparatively precise adenocarcinoma ALK+ status diagnosis automatically, meanwhile objectively.

Lately, a handful of radiogenomics studies on lung cancer involved contrast CT images: Tu et al²⁰ built an integrated EGFR+ NSCLC model (training cohort AUC/validation cohort AUC = 0.798/0.818) with pre-contrast CT radiomics, clinical and morphological data, which outperformed a clinical model. Zhao et al developed a deep learning system based on 3D convolutional neural network (CNN) to predict EGFR+ in pulmonary adenocarcinoma and achieved an AUC of 75%.²² He et al⁴⁰ suggested that early contrast enhancement within 25 s affected the ability to diagnose lung nodules, lung cancer, and that non-contrast CT images showed better diagnostic performance. However, in clinical practice, a 25 second delay, which could be too early for lung lesion enhancement to show up, is not the standard timing for lung contrast enhancement. Kakino et al⁴¹ claimed that contrast enhancement in the delayed phase (120 s) of CT image for NSCLC patients affected some of the radiomics features and the variability of radiomics features, compared with non-contrast CT images. Of note, the performance of this radiogenomics classier was improved by replacing pre-contrast CT images with clinical standard post 38–48 s contrast CT images. One possible reason for this could be that the contrast enhancement during a proper time window may bring forward additional information that could be utilized by AI, suggesting that proper post-contrast CT radiomics should be recommended in the context of future NSCLC radiogenomics AI approaches.

In this study, two different scanners from the same manufacturer were adopted. A previous study⁴² evaluated NSCLC images from 100 CT scanners (including different manufacturers) at 35 clinics and found the radiomics varieties, which can be significantly reduced by image normalization and standardization,⁴³ were relatively small. Hence, before analysis, we applied standardization and normalization to all the matrices. According to a recent NSCLC EGFR radiomics study, compared with thick CT slice (5 mm), thin slice (1 mm) achieved the best prediction value, independently of convolution kernels.⁴⁴ In this retrospective study, nevertheless, in view of image standardization among each scan and avoiding interpolation, all CT images carried equal slice thickness of 5 mm.

In conclusion, this preliminary radiogenomics machine learning suggested: 1) solid lung adenocarcinoma ALK+ radiogenomics classifier of standard post-contrast CT radiomics biomarkers produced superior performance compared with that of pre-contrast one, suggesting that post-contrast CT radiomics should be recommended in the context of lung adenocarcinoma radiogenomics AI. 2) Standard post-contrast CT machine learning radiogenomics classifier could help to precisely identify solid lung adenocarcinoma ALK rearrangement status, which may act as a pragmatic and cost-efficient substitute for traditional invasive ALK status testing.

This study is limited by relatively small sample size, which cannot reflect the overall ALK+ adenocarcinoma population. Additionally, selection bias is inevitable due to its retrospective nature. Also, only three leading categories of adenocarcinoma were involved in this study; some rare genomic categories with low prevalence rate were not included. Therefore, a further prospective multicenter whole-genome adenocarcinoma radiogenomics study is called for to validate our current results.