CT-Based Radiomics Can Predict the Efficacy of Anlotinib in Advanced Non-Small-Cell Lung Cancer

Anlotinib is a small-molecule RTK inhibitor that has achieved certain results in further-line treatment, but many patients do not respond to this drug and lack effective methods for identification. Although radiomics has been widely used in lung cancer, very few studies have been conducted in the field of antiangiogenic drugs. This study aims to develop a new model to predict the efficacy of patients receiving anlotinib by combining pretreatment computed tomography (CT) radiomic characters with clinical characters, in order to assist precision medicine of pulmonary cancer. 254 patients from seven institutions were involved in the study. Lesions were selected according to the RECIST 1.1 criteria, and the corresponding radiomic features were obtained. We constructed prediction models based on clinical, NCE-CT, and CE-CT radiomic features, respectively, and evaluated the prediction performance of the models for training sets, internal validation sets, and external validation sets. In the RAD score only model, the area under curve(AUC) of the NCE-CT cohort was 0.740 (95% CI: 0.622, 0.857) for the training set, 0.711 (95% CI: 0.480, 0.942) for the internal validation set, and 0.633(95% CI: 0.479, 0.787) for the external validation set, while that of the CE-CT cohort was 0.815 (95% CI: 0.705, 0.926) for the training set, 0.771 (95% CI: 0.539, 1.000) for the internal validation set, and 0.701 (95% CI: 0.489, 0.913) for the external validation set. In the RAD score-combined model, the AUC of the NCE-CT cohort was 0.796 (95% CI: 0.691, 0.901) for the training set, 0.579 (95% CI: 0.309, 0.848) for the internal validation set, and 0.590 (95% CI: 0.427, 0.753) for the external validation set, while that of the CE-CT cohort was 0.902 (95% CI: 0.828, 0.977) for the training set, 0.865 (95% CI: 0.696, 1.000) for the internal validation set, and 0.837 (95% CI: 0.682, 0.992) for the external validation set. In conclusion, radiomics has accurate predictions for the efficacy of anlotinib. CE-CT-based radiomic models have the best predictive potential in predicting the efficacy of anlotinib, and model predictions become better when they are combined with clinical characteristics.


Introduction
Lung cancer is the malignant tumor with the highest mortality rate in the world. In 2020, the incidence rate of lung carcinoma was 11.4% and the mortality rate was as high as 18% [1]. With the advent of targeted therapy and immunotherapy, treatment options for lung cancer patients have become more diverse, and the overall survival also continues to prolong. Te 3-year survival rate rose from 19% in 2001 to 31% in 2015 through 2017, and the median survival increased from 8 to 13 months, with the fve-year survival rate for non-small-cell lung cancer (NSCLC) being higher than that for small cell lung cancer [2,3]. As treatment options become more available, how to make a patient's treatment decision remains controversial.
Receptor tyrosine kinases (RTKs) are single-pass membrane proteins closely related to cell growth, motility, diferentiation, and survival. Tey are grouped into subfamilies based on the similarity of their extracellular domains, including vascular endothelial growth factor (VEGF) or fbroblast growth factor receptor (FGFR) families [4]. Types of tyrosine kinase inhibitors (TKIs) demonstrate efects on inhibition of angiogenesis; however, only a few TKIs, including nintedanib and anlotinib, have shown positive anticancer efects [5]. Anlotinib is a small-molecule RTK inhibitor that targets VEGFR1, VEGFR2/KDR, VEGFR3, and Raf serine/threonine kinases, platelet-derived growth factor receptor PDGFR-α, and FGFR1, FGFR2, and FGFR3 [6,7]. Terefore, anlotinib can inhibit angiogenesis in tumors and limit tumor growth. ALTER-0303 was a randomized, double-blind, placebo-controlled, multicenter, phase III trial that compared the efcacy and safety of anlotinib with that of the placebo in patients with advanced NSCLC who progressed after at least two lines of prior treatments. Te results showed that anlotinib was more efective than the placebo in third-line treatment in patients with advanced NSCLC [8]. Tough anlotinib has achieved certain results in further-line treatment, there are still many patients who do not respond to this drug. How to fltrate patients efectively for anlotinib remains unclear.
Diferent from traditional needle biopsy methods to obtain local tissue samples, radiographic images can fully display the overall characteristics of tumors to analyze the heterogeneity of tumors. Based on machine learning, radiomics has been widely used in the diagnosis and treatment of oncology. At present, many studies have used radiomics to detect the pathological type, gene mutation, programmed death ligand 1 (PD-L1) expression level, or tumor-infltrating lymphocyte (TILs) levels in patients and even predict the efcacy of types of treatment [9][10][11][12][13][14][15]. A meta-analysis showed that the pooled diagnostic odds ratio for predicting immunotherapy response in NSCLC using radiomics was 14.99 (95% CI: 8.66-25.95) [16]. Tese research studies demonstrate the potential of radiomics for solving clinical problems.
Although radiomics has been widely used in lung cancer, very few studies have been conducted in the feld of antiangiogenic drugs. Tis study aims to develop a new model to predict the efcacy of patients receiving anlotinib by combining pretreatment computed tomography (CT) radiomic characters with clinical characters, in order to assist precision medicine of pulmonary cancer. In this study, patients who underwent noncontrast-enhanced CT (NCE-CT) and CE-CT were divided into two groups for feature extraction and training separately [17].

Patient Population.
Te data on the training set and the internal validation set was retrieved from a clinical trial: "A Real-world Study: Efcacy and Safety of Anlotinib for Advanced Non-small Cell Lung Cancer (NSCLC)" (NCT04871997). Te data on the external set was obtained from the medical record system of Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology, and these patients did not participate in the clinical trial. In this retrospective study, all the patients had been getting recruited since July 2019, ending in September 2022. Data had been getting analyzed since October 2022. Tis study was approved by the Institutional Ethics Committee of Tongji Medical College, Huazhong University of Science and Technology (approval number S1040). Informed consent was obtained from all patients.
Inclusion criteria were as follows: (1) age: ≥18 years old, no gender limit; (2) diagnosed as advanced nonsmall-cell lung cancer; (3) at least one tumor lesion in the lung was not subjected to local treatment such as irradiation in the past and could be accurately measured, and the longest diameter was ≥10 mm; (4) the patient received two cycles of medication and at least had efcacy assessment performed once. (5) Patients included in the training set needed to achieve the median progression-free survival (PFS). If any of the above items was "no," the patient was not suitable for this study. Exclusion criteria were as follows: (1) those who were confrmed to be allergic to anlotinib and/or its excipients; (2) patients with anlotinib contraindications. Finally, 254 patients were retrospectively involved in this study. Table S1, all CT images were acquired by using a multislice CT system, with a tube voltage of 120 kVp and automatic tube current modulation. Each CT image was reconstructed with an image matrix of 512 × 512 pixels, while the slice thickness of CT scans was 1.0 to 2.0 mm with coverage from the apex to the bottom of the lungs. Images were saved in the DICOM format.

Region-of-Interest Segmentation and Feature Extraction.
Tumor lessons were segmented by a chest radiology specialist with 1 year of experience using open-source software, 3D-slicer (version 4.11, http://www.slicer.org). A senior imaging physician (LJL with 7 years of experience) performed the identifcation of region-of-interest (ROI) regions. Intrapulmonary lesions were selected as per the RECIST version 1.1 criteria [18], and for patients possessing multiple evaluable lesions in the lungs, the fve lesions with the largest diameter were selected for outlining. Using the "threshold" semiautomatic segmentation tool in the range of WW −500∼−100 HU and WL 200-400 HU to segment the lesions in the lungs confrmed the extent of the tumor and excluded vascular and gas shadows within the tumor. Te average radiomic features of all target lesions were used as global radiomic features to predict treatment response.
A total of 851 original radiomic features were extracted using a PyRadiomics (version 3.0.1: Computational Imaging and Bioinformatics Lab, Harvard Medical School) module in 3D-slicer from these ROIs, consisting of frst-order features (N � 18), shape features (N � 14), and texture features (N � 75) extracted from preprocessed original images, and wavelet features (N � 744) were extracted from processed images with wavelet flters. Te category of features is displayed in Table S2.

Radiomic Feature Selection and Radiomic Signature
Development. Tis study complied with the IBSI guidelines in general. To avoid the information loss of processing, we did not perform any special pretreatment for radiographic imaging considering that the diference in CT parameters in this study were not particularly large [19]. After obtaining image omics parameters, the t-test was conducted for each omics parameter based on the corresponding prognosis. Parameters with P < 0.1 were screened for subsequent LASSO logistic modeling. LASSO logistic regression was used for dimension reduction and identifying the most PFSrelated features from repeatable and nonredundant features in the training set. Patients of NCE-CT and CE-CT were randomly divided into training and validation sets at the ratio of 3 : 1. Te radiomic score (RAD score) was calculated through a linear combination of selected features weighted by their respective coefcients.

Primary Outcome Measures
(1) Progression-Free Survival (PFS). It is the length of time during and after the treatment of a disease, such as cancer, that a patient lives with the disease, but it does not get worse.

Secondary Outcome Measures
(1) Overall Survival (OS). It is the length of time from either the date of diagnosis or the start of treatment for a disease, such as cancer, that patients diagnosed with the disease are still alive.
(2) Objective Response Rate (ORR). It is the percentage of patients on whom a therapy has some defned efect; for example, cancer shrinks or disappears after treatment.
(3) Disease Control Rate (DCR). It is the percentage of patients with advanced or metastatic cancer who have achieved complete response, partial response, and stable disease with a therapeutic intervention in clinical trials of anticancer agents.

Patient Population.
In this study, complete treatment history and follow-up data were collected from 362 patients, and after censoring 38 patients who were unable to obtain baseline CT, 28 patients who lacked intrapulmonary lesions, and 14 patients who did not obtain imaging efcacy assessment, 282 patients were included in the study. CT was performed within 14 days prior to drug administration. Patients included in the model training set were required to achieve the median PFS or disease progression, so 28 patients were excluded. Finally, 254 patients were included in the study. Te participants' fowchart is displayed in Figure 1.
Baseline patient characteristics are summarized in Table 1. Tere was no signifcant diference in the training set, the internal validation set, and the external validation set in terms of sex, age, smoking, Eastern Cooperative Oncology Group Performance Status (ECOG PS) score, metastasis lessons, pathological type, treatment line, and best response.

Acquisition of Radiomic Features and Construction of Models.
A total of 851 radiomic features were extracted from CT images for each patient. Tese extracted features were preprocessed using t-tests to remove redundant and irrelevant features. Ten, LASSO regression modifed 7 features for the NCE-CTcohort and 5 features for the CE-CT cohort, which were mostly associated with PFS in the training set. Te RAD score was calculated through a linear combination of selected features weighted by their respective coefcients (Appendix S1). Te training process is shown in Figure 2.
It can be observed that NCE-CT contains one parameter related to tumor size and that all other parameters are texture features, while the CE-CT parameters are all texture features. CE-CT is able to better visualize microvascular changes within the tumor due to the presence of contrast agents, thus enabling a better characterization of tumor heterogeneity, which in turn has better potential for predicting the efcacy of anlotinib.

Predictive Validity of the Model.
After performing logistic regression for the NCE-CT cohort and the CE-CT cohort separately, the RAD score was used to set up a model in conjunction with the logistic regression results, and we also performed modeling using the RAD score alone to obtain a total of four ROC charts and two nomogram plots It can be clearly seen that for the model with the RAD score alone, the NCE cohort exhibited less than optimal predictive validity, while the CE cohort showed some predictive power. After adding clinical factors to the model, the predictive power of the NCE cohort decreased, while the predictive power of the CE cohort signifcantly improved.
DCA was performed for the above two radiomic only models as well as the model built based on radiomics and clinical features. As shown in Figure 4, the training sets of all four models have good DCA fts, while the validation sets of the combination of the radiomic and clinical feature model and the CE-CT radiomic only model have better DCA performance. Te NCE-CT radiomic only model has relatively poor results, which also coincide with the results of the ROC curve.

Verifcation of the Predictive Power of the Model in the External Validation Set.
Te Youden index in the ROC curve was used as our threshold for diferentiating patients into high-risk and low-risk populations (Figures 3(b) and 3(e)) represent the radiomic only models, and Figures 3(c) and 3(f ) represent the radiomic-combined models). Patients with logistic scores below this threshold were in the high-risk group, and vice versa in the low-risk group, and the PFS showed a signifcant diference in the internal and external validation set. As shown in Figure  In terms of OS, low-risk patients were signifcantly higher than high-risk patients in the internal validation set of the radiomic only model (Figure 6(a), P � 0.0008). For the external validation set, the radiomic only model and in both sets of the radiomic-combined model, the diference in OS was not signifcant between the high-risk and low-risk populations, but we can still see certain trends through the KM curves (   prognosis of the high-risk and low-risk cohorts for either PFS or OS (Figures S1 and S2).

Discussion
Radiomics has been widely used in the whole process of cancer treatment, among which the clinical application of tumor antiangiogenic drugs and imaging in lung cancer has been very extensive. Most of the studies on antitumor angiogenic drugs are focused on bevacizumab, and most of the tumor types are recurrent glioma and gastrointestinal tumors. Several studies have used MRI-based radiomics to predict the efcacy of bevacizumab in patients with recurrent glioma, and excellent predictive validity has been observed for PFS and OS [20,21]. In gastrointestinal tumors, CE-CT radiomic studies of patients with colon and liver cancers were successful in predicting the efcacy of bevacizumab in patients receiving bevacizumab [22,23]. Compared to the above studies, radiomic studies targeting small-molecule antitumor angiogenic agents are very rare. PET, CT, and MRI-based radiomics in kidney cancer can efectively predict early response and survival with sunitinib but mostly in small sample  36  35  35  35  35  35  34  35  30  28  23  23  23  22  20  18  15   Journal of Oncology    studies [24][25][26]. Another small sample study explored the efcacy of combined CT and methotrexate-based prediction of apatinib for advanced hepatocellular carcinoma with favorable results [27]. Despite the scarcity of such studies, they show us the good potential of radiomics for the prediction of efcacy of antitumor angiogenic drugs.
Te study of predicting the efcacy of antitumor angiogenic drugs in lung cancer by radiomics is scarce. In our study, we built models to predict the efcacy of antiangiogenic drugs based on NCE-CT and CE-CT, respectively. Te radiomic features incorporated into the modeling are predominantly second-order features, implying that texture features act as primary predictive correlates, and we found that the models constructed using CE-CT images had good predictive validity and were more efective when clinical factors were added. Te NCE-CTbased model did not have such good predictive efcacy. Tere was one study using both CE-CT and NCE-CT images to model the predicted prognosis of immunotherapy, but there was no signifcant diference in the predictive performance of the two models [28]. In another study, using CE-CT and NCE-CT to predict EGFR mutation status in NSCLC patients, the predictive performance of the two methods also did not difer signifcantly [29]. Anlotinib acts as an antiangiogenic agent, which primarily afects the tumor microvasculature that has an efect on the tumor. Due to the presence of contrast agents, CE-CT is able to better visualize microvascular changes within the tumor. Tis may be the reason why the model based on CE-CT in our article was able to better predict the efcacy of anlotinib.
Tis article, as the frst article exploring NCE-CT versus CE-CT in the prediction of anlotinib and with patients derived from clinical studies, has high credibility of information and bright results but still leaves some questions. (1) Heterogeneity of imaging parameters resulting from multicenter studies may afect study results because no resampling was performed. However, there is not much diference in parameters between individual CT instruments. We also performed an external validation and confrmed that our model has good stability and accuracy. (2) Heterogeneity between the training and validation sets might exist after grouping due to insufcient number of patients. (3) Anlotinib, as a drug developed in China, data are only available for the Chinese population. (4) Tis is a real-world study, and patients may be combined with other treatment options during the course of their treatment, which may afect the accuracy of the results.

Conclusions
Anlotinib has good efcacy in the treatment of advanced non-small-cell lung cancer. Radiomics has accurate predictions in the efcacy of anlotinib. CE-CT-based radiomic models have the best predictive potential in predicting the efcacy of anlotinib, and model predictions become better when they are combined with clinical characteristics.

Data Availability
Te raw data and source codes used to support the fndings of this study have been deposited in the GitHub repository (https://github.com/ChuhuaiWang/data_analysis_code).

Disclosure
Jingyu Chen and Chuhuai Wang are the co-frst authors.