Introduction
Multimodal medical images contain information invisible to the examiner’s eye that can provide important data about the diagnosis, prognosis, and course of cancer. The advances in machine learning make it possible to extract many quantitative features and transform the medical images into data. The method, first proposed in 2012, which consists of extracting and analyzing data from medical imaging in order to improve the accuracy of diagnosis and treatment, has been called radiomics. The concept is based on extracting specific features from computed tomography (CT) imaging, magnetic resonance imaging (MRI), positron emission tomography (PET), ultrasonography, digital radiography (DR) and subsequently correlating them with biological and clinical data in order to build models that improve accuracy and the performance of the diagnosis and to estimate before applying a treatment the possible evolution of the patient to the therapy. The number of scientific papers evaluating the capacity of radiomic techniques in medicine is constantly growing and oncology is one of the fields in which radiomic analysis opens new horizons(1-3).
Although the potential of this method to be used as a biomarker in oncology is huge, the standardization of data acquisition methods is a necessity for the method to be implemented in the oncology clinical routine to allow the monitoring of tumor and tumor microenvironment in order to predict the individualized prognosis or to anticipate the evolution and response to therapy(4).
Lung cancer is the leading cause of cancer mortality worldwide in both men and women, accounting for approximately 25% of cancer deaths, with late diagnosis being probably the most significant factor for this unfavorable prognosis. To date, no effective and noninvasive screening method has been identified in lung cancer, and in this context the potential of radiomic analysis in this direction is considerable. Non-small cell lung cancer (NSCLC) includes a variety of cancers, the most common being squamous cell carcinoma and adenocarcinoma. Squamous cell carcinoma was considered the most common histological type, being associated with smoking, but with the implementation of anti-tobacco campaigns, there was an increase in the incidence of adenocarcinoma. Radiomic analysis demonstrates the possibility to facilitate the differential diagnosis of a benign and a malignant nodule, to place the patient in a prognostic group, to predict the histological and molecular type, and to predict the response of each case to oncological therapy(5-6).
Radiomics and radiogenomics –
from translational lung cancer research
to clinical practice
New molecular and genetic analysis techniques have led to the division of cancers into molecular subtypes, with direct implications in prognosis and in the identification of new therapeutic targets, the mutational status of EGFR, KRAS and TP53 being recently considered the most significant in establishing the prognosis of adenocarcinoma. Radiomics and especially its “sister”, called radiogenomics, allow the noninvasive and rapid exploration of molecular subtypes, correlating mining data from images with mutational status, in order to create machine learning models to obtain an accurate molecular diagnosis before the analysis of the pathological sample. The algorithm of the radiomechanical analysis method has several stages, the first one starting with the acquisition of the image. In performing a radiomic analysis, for an accurate segmentation of the tumor, the second stage of the radiomic process requires a more accurate standardization of the acquisition parameters. Once the region of interest (ROI) is defined, it can be the whole tumor, a part of it, or the tumor and the microenvironment. The next step of the analysis is to extract the radiomic characteristics using software applications. With the increase in resolution and software development, the number of radio features extracted from medical images has increased. Among the categories of radiomics features that are used in radiomic analysis there are morphological characteristics that reflect the shape and physical characteristics of ROI, histogram-based features that do not retain spatial information, and texture features based on pixel neighborhood information. These features take into account the pixel and its neighbors, the gray level co-occurrence matrix (GLCM) being one of the most used features in this category. The selection of characteristics and the construction of a model follow, using the most significant radiomic characteristics with prognostic or predictive value in the desired clinical application(7-11).
The prognosis of the disease and the overall survival (OS) were analyzed in radiomics studies by Grove and collaborators, who identified parameters of heterogeneity such as spicularity and entropy gradients as prognostic indicators in patients with early-stage lung cancer. Other authors have identified textural features related to survival and local control, but these features may also be associated with the risk of metastases. A total of 91 stage III NSCLC cases treated with definitive radiochemotherapy, imagistically evaluated by pre-treatment CT, followed by radiotherapy planning simulation with 4-dimensional CT (4D-CT), were included in a study that identified textural features such as histogram, gradient and GLCM as having a signifiant prognostic value. The granulomatous nodule with spicular or lobulated aspect generate problems of differential diagnosis with solid pulmonary adenocarcinoma, based on the morphological and metabolic features evaluated from CT and positron emission tomography (PET)/CT imaging, being reported high rates of false-positive diagnoses. The aim of a study was to investigate the ability of radiomics using data extracted from CT images to make a differential diagnosis of accuracy. The study included 302 patients who were evaluated, the patients’ lot being divided into the training cohort and the validation cohort, including 211 and 91 patients, respectively. The radiomic features were extracted from CT images using a computer software application. A logistic regression model was used to identify the optimal radiomic characteristics and, subsequently, a radiomic model was constructed for the differential diagnosis of the granulomatous nodule of adenocarcinoma. Previously analyzed data reported rates of 16.7% of adenocarcinomas that were misdiagnosed as granulomatous nodule and 24.7% granulomatous nodules were misdiagnosed as bronchopulmonary neoplasm before surgery. The radiomic evaluation associated with the multivariable model, including clinical risk factors, had better results than the radiomic models used as the only method, the results being similar in case of image evaluation by expert imagers and in case of radiomic evaluation(12-15).
Another retrospective study evaluating the ability of the radiomics method to perform the differential diagnosis between benign/malignant in lung cancer used features extracted from CT images for 290 patients and, subsequently, the histopathological analysis was performed for the suspected nodules. Node shape, Gabor function and texture-based characteristics (energy of Haralick and Laws energy) were extracted from images, using intranodal and perinodular regions for features extraction. For a study including 145 patients, the results of the radiomic classifier were compared with a convolutional neural network and the diagnostic assessment of two imaging experts. The combination of the radiomic detection both for intranodular and perinodular regions improved the correct diagnosis rate, exceeding the accuracy of the differential diagnosis offered by the subjective evaluation of the imaging experts. Another radiomic analysis aims to evaluate the predictive value of texture information obtained from nodular opacities identified from CT imaging, having as objective the evaluation of the capacity of the method to predict the recurrence of adenocarcinoma. The prediction rate seems to be significantly higher when the attributes of the images were based on the solid components of the tumors compared to the case when the entirely segmented tumors were analyzed. This radiomic approach suggested that most information on adenocarcinoma related to the disease aggression is related to density and the morphological properties of solid tumor subvolumes. The model also evaluated the ability of radiomics to predict adenocarcinoma recurrence, the method having in this case a low specificity, but a very high sensitivity. Radiomics also demonstrates the ability to identify patients at low risk for post-surgical recurrence and to accurately estimate the risk of recurrence at a certain time after surgical resection(16,17).
Radiomics also offers the possibility to predict distant metastases risk for patients with pulmonary adenocarcinoma, using the analysis of radiomic features extracted from CT imaging. For this purpose, Coroller et al. used two patients lots (a lot including 98 patients for radiomic model training, and a lot with 84 cases for the model validation). Primary tumor phenotype was quantified on pre-treatment CT scans, using 635 radiomics features, and a univariate and a multivariate analysis was performed to assess the radiomic model performance based on the concordance index concept. The authors identified 35 radiomics features that were found to be prognostic for distant metastases and 12 features considered correlated with patients’ survival. Tumor volume was only a moderate prognosis imagistic biomarker for distant metastases. The addition of the radiomic signature to a clinical model resulted in a significant improvement in the power to predict the presence of distant metastases in the validation set(18).
Radiomatics also demonstrated the ability to predict the pre-treatment pathological response after neoadjuvant chemoradiation in patients with NSCLC. A study including 127 cases used 15 radiomic features but also classic features such as tumor volume and diameter. Seven of the radiomic features were considered predictive for the residual disease and one for the complete pathological response, but the volume and diameter of the tumor were not identified as having predictive value. However, the round shape and the heterogeneous texture were associated with an unfavorable response to chemotherapy. The study demonstrated the ability of radiomic analysis to have superior predictive power over conventional imaging tumor features(19).
Radiomics was investigated to identify CT imaging predictors of clinical outcomes in early NSCLC patients treated with stereotactic body radiotherapy (SBRT). Analyzing the CT images from 113 NSCLC stage I-II patients treated with SBRT from a radiomic point of view, the authors chose 12 radiomic features that were also compared with conventional imaging parameters (tumor volume and diameter) and with clinical parameters. OS was associated with tumor volume and diameter, and with two radiomic features. One radiomic feature was significantly prognostic for distant metastases and, also, it should be noted that none of the conventional and clinical parameters were significant to assess the metastatic risk(20).
Tumor mutation burden (TMB) is known to be a significant predictor for the effectiveness of immunotherapy in many cancers that benefit from immune checkpoint inhibitor (ICI) therapy. The modern methods of deep learning in radiomic analysis have been investigated for the evaluation of radiomics as a biomarker of response to ICI treatment in patients with NSCLC. In a radiomic study, He et al. included the CT images from 327 patients with a known TMB data. The radiomic biomarker was able to differentiate in a validation cohort, after the model training, the patients with low and high TMB in the test cohort. Radiomic deep learning methods thus demonstrate the ability to predict the response to ICI therapy in advanced NSCLC, and the authors mention that the inclusion of Eastern Cooperative Oncology Group (ECOG) performance status and the tumor microenvironment assessment increase the predictive power of radiomic analysis(21).
Conclusions
Radiomics, a new subdomain of artificial intelligence, has demonstrated promising prospects in creating prognostic and predictive models in oncology. The use of radiomics and, more recently, of the deep learning algorithms in the management of lung cancer demonstrates the possibility of an accurate differential diagnosis between lung cancer and benign nodules, but can also predict the risk of recurrence and metastasis at a certain interval after surgery. The radiomic biomarker also demonstrates the ability to predict the response to chemoradiotherapy, SBRT or to immunotherapy. Radiogenomics – the radiomics “sister” – demonstrates the ability to predict noninvasively the noninvasive molecular subtypes of lung cancer, many of these subtypes benefiting from molecular therapeutic targets. After the issue of standardization is overcome, radiomics can be included in the clinical decision algorithm, having the advantage of being a fast, low cost and noninvasive method.
Conflicts of interests: The authors declare no conflict of interests.