Potential Application of Radiomics for Differentiating Solitary Pulmonary Nodules

Kaikai Wei; Huifang Su; Guofeng Zhou; Rong Zhang; Peiqiang Cai; Yi Fan; Chuanmiao Xie; Baowei Fei; Zhenfeng Zhang

doi:10.4172/2167-7964.1000218

ISSN:2167-7964

OMICS Journal of Radiology

Make the best use of Scientific Research and information from our 700+ peer reviewed, Open Access Journals that operates with the help of 50,000+ Editorial Board Members and esteemed reviewers and 1000+ Scientific associations in Medical, Clinical, Pharmaceutical, Engineering, Technology and Management Fields.

Meet Inspiring Speakers and Experts at our 3000+ Global Conferenceseries Events with over 600+ Conferences, 1200+ Symposiums and 1200+ Workshops on Medical, Pharma, Engineering, Science, Technology and Business

Potential Application of Radiomics for Differentiating Solitary Pulmonary Nodules

Kaikai Wei^1#, Huifang Su^1#, Guofeng Zhou^2#, Rong Zhang¹, Peiqiang Cai¹, Yi Fan³, Chuanmiao Xie^1, Baowei Fei^4 and Zhenfeng Zhang^1*
¹Sun Yat-sen University Cancer Center, State Key Laboratory of Oncology in South China, Collaborative Innovation Center for Cancer Medicine, Guangzhou, People’s Republic of China
²Department of Radiology, Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, People's Republic of China
³Department of Radiation Oncology, Perelman School of Medicine, University of Pennsylvania, Philadelphia, USA
⁴Department of Radiology and Imaging Sciences, Emory University School of Medicine; Department of Biomedical Engineering, Emory University and Georgia Institute of Technology, Atlanta, USA
^#These authors contributed equally to this work.
*Corresponding Authors:	Chuanmiao Xie Sun Yat-sen University Cancer Center State Key Laboratory of Oncology in South China Collaborative Innovation Center for Cancer Medicine Guangzhou, People’s Republic of China E-mail: xiechm@sysucc.org.cn
	Baowei Fei Department of Radiology and Imaging Sciences Emory University School of Medicine Department of Biomedical Engineering Emory University and Georgia Institute of Technology, Atlanta, USA E-mail: bfei@emory.edu
	Zhenfeng Zhang Sun Yat-sen University Cancer Center State Key Laboratory of Oncology in South China Collaborative Innovation Center for Cancer Medicine Guangzhou, People’s Republic of Chinas Tel: +86 0208 7342 119 E-mail: zhangzhf@sysucc.org.cn
Received date: December 08, 2015; Accepted date: March 18, 2016; Published date: March 21, 2016
Citation: Wei K, Su H, Zhou G, Zhang R, Cai P, et al. (2016) Potential Application of Radiomics for Differentiating Solitary Pulmonary Nodules. OMICS J Radiol 5:218. doi:10.4172/2167-7964.1000218
Copyright: © 2016 Wei K, et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Visit for more related articles at OMICS Journal of Radiology

View PDF Download PDF Tables & Figures

Abstract

A solitary pulmonary nodule is defined as radiographic lesion with diameters no more than 3 cm and completely surrounded by normal lung tissue. It is commonly encountered in clinical practice and its diagnosis is a big challenge. Medical imaging, as a non-invasive approach, plays a crucial role in the diagnosis of solitary pulmonary nodules since the potential morbidity of surgery and the limits of biopsy. Advanced hardware, image acquisition and analysis technologies have led to the utilization of imaging towards quantitative imaging. With the aim of mining more useful information from image data, radiomics with high-throughput extraction can play a useful role. This article is to introduce the current state of radiomics studies and describe the general procedures. Another objective of this paper is to discover the feasibility and potential of radiomics methods on differentiating solitary pulmonary nodules and to look into the future direction of radiomics in this area.

Keywords

Radiomics; Pulmonary nodules; Differentiation

Introduction

The term “Radiomics” derives from the combination of word “radiology” and the suffix “omics”, and it is a new extension of omics method applied in radiology quantitative study. Radiomics refers to extracting and analyzing features from medical images systemically [1]. Segal and his colleagues had brought out one of the pioneer study on radiomics, which indicates the strong relationship between CT (computed tomography) image traits and gene expression in liver cancer (78% global gene expression profiles were constructed by combination of 28 imaging traits) [2]. Extracting and analyzing approaches of image data in this article give us the sight of deeply interpretation of image sources. However, this is a retrospective study and the sample size is small, besides, the reliability of imaging traits were all manually assessed, all these made an inevitable bias in this research. As the consequence of development of image segmentation and statistical analysis, the reproducibility and stability of delineation of lesion are higher and higher, also more and more analysis pattern are developed, all these allow the quantitative imaging to move towards radiomics, where a lot of features are extracted and interpreted. A study concerning 1019 patients has proved the correlation between radiomic signature and gene expression in lung or head and neck cancer, suggesting a cost effective access in cancer treatment [3].

Solitary Pulmonary nodule is a common encountered issue in clinical practice. It has been defined as “a rounded or irregular opacity, well or poorly defined measuring up to 3 cm in diameter” on CT image [4]. Pre-test possibilities of malignant act a pivotal part in management of solitary pulmonary nodules. Clinical factors, radiologic features and biopsy are commonly used in distinguishing malignant from benign of solitary nodules. Clinical factors such as location in the upper lobe were confirmed to be increased possibility of lung cancer in an evidence-based prospective study [5], however, some clinical history show its limit use, for instance, smoking history are no longer a factor in determination of malignancy of pulmonary nodules since increasing population of lung adenocarcinoma in younger and non-smoking patients [6]. Biopsy is considered to be a “golden standard” for insight into possible cancerous, but as invasive procedures, the applications of biopsy are limited, for example, the complication after sampling. Radiologic features as non-invasive approaches is most commonly used in clinical practice, features like margin, size, calcification are assessed by radiologists and then integrated into an inclination of malignancy. Accuracy of diagnosis boosts owing to the innovation of new scanners, image agents and standardized protocols, which allow radiologists obtain more useful features and thus make a more accurate diagnose. It can be concluded that the more useful features of lesions we can achieve, the more accurate diagnosis we are able to make. In this paper, we are going to introduce the radiomics study and argue the feasibility and potential of radiomics applied in differentiating solitary pulmonary nodules (Figure 1).

Hardware and Software for Radiomics Study

Original data and analysis methodology are two core elements of radiomics study that serves as hardware and software of radiomics. Data mainly contains image data and clinical data while analysis tools generally concerning extracting and statistical analyzing. One destination of radiomics study is trying to give a prediction of prognosis or a guidance of treatment pilot. The precondition to achieve this goal is that stable methods of feature obtained. These methods ought to be characterized with reproducible, feasible, stable, and these characteristics should runs through lesion delineation and feature extraction. Research including 32 patients has found that 29 features are reproducible among 329 image features [7], another study with 56 patients concerning the reproducibility of quantitative image features among 3 different CT machine has been conducted, and the result indicates that 138 out of 328 features are reproducible [8]. Except for features, the assessing tools are also considered to be an important part of radiomics studies, since an objective accurate evaluating method make the result more convinced. AUC (Area under Curve) of the ROC (Receiver Operating Characteristic Curve) and CI (Concordance Index) have been used in assessing the reproducibility of the feature and the performance of models in the meaning of interpreting the information the features take. These results indicate the potential of radiomics features being applied clinically. However, these results still needs further demonstrated with large sample study and various assessing methods.

Images from different modalities are served as data source or data pool, they are later imported into software, which segments and then exports. The series of software mentioned above allow operators to segment the images manually, semi-automatically or automatically, and then export region of interest (ROI) of images for further digital feature analysis [9,10]. Large amounts of features have been studied, and the number of feathers is still increasing. These features are expressed as numerical patterns, and biological information is considered to be concealed behind them. For example, texture feature, defined as a function of the spatial variation in pixel intensities, which has been widely used in computer science, is now applied in radiomics. Features need to be independent, informative, stable, and reproductive, and these crucial characteristics make them the corner stones in building either prediction or diagnosis models, and yet some of them are indicated to be workable candidates with high reproducibility for accessing tumor and biomarker [7,11-13]. Meanwhile, modern algorithms allow us to handle these large amounts of features and select the most stable, reproducible, independent, and informative ones among them. For example, mRMR (minimum Redundancy Maximum Relevance) feature selection is an algorithm that gained considerable improvement in feature selection and classification accuracy, which were later used in radiomics research [14,15]. Present studies are trying to discover as many features as possible; extensive features and optimal algorithms play a foundational role in the development of radiomics. To achieve this goal, it is necessary to build a good collaboration between medical experts and computer engineers.

Radiomics study of Cancer

One pioneer’s attempt was put into practice in 2007 when researchers tried to decode gene expression programs by exploring radiographic features in liver cancer [2]. The trial earned a positive result, which showed that the gene expression profiles of hepatic carcinoma were associated with their radiographic features suggesting that the features extracted from images were able to reflect gene expressions in a convoluted manner. Studies of radiomics on cancer have sprung up after the very first attempt of radiomics on liver cancer. These studies was first focused on molecular aspect trying to decode gene expression by radiomics, and now more and more territory of both experimental and clinical medicine, for preclinical studies, the relationship between gene patterns or expression and imaging are investigated, for clinical medicine, various clinical study interest are raised.

Medical images are collected retrospectively and investigated with patients’ information as well as follow-up data. These studies are actually to demonstrate the relationship between imaging features and information with directly or indirectly significant benefits for diagnosis or therapy, and then to develop objective and robust features as descriptors that can be used in clinic. Various features have been validated to be associated with biological information ranging from genes to tissue in many diseases. For instance, image features were demonstrated to reveal the association with gene expression either in level or patterns in liver cancer, lung cancer, head and neck cancer [2,16]. The same results also suggested that the relationship between features and heterogeneity in lung and liver cancers [17]. Those microinvisible characteristics may directly associate with overall survival, disease development and responses to treatments. It has been demonstrated that overall survival was related closely with image features in lung, breast, and colorectal cancer [18-20]. Recently it was attested that some imaging features strongly predict the distant metastasis of adenocarcinoma after radiotherapy [14], and radiomicsbased features were also studied to be associated with pneumonitis development or metastasis after radiotherapy [21], whereas most of these researches are retrospective and small sample sized, more evidence should be provided by large scale perspective studies.

As for response to therapies, most studies focused on drug response while in fact there are more we can do. For instance, radiotherapy dose can be linked with radiomics-based features [21,22]. All these indicated the potential ability of radiomics feather as a biomarker in predicting and surveillance of diseases; some scientists had pioneer studies on it and obtained encouraging results [23]. The extensive application of radiomics prospects a bright future. Besides, we also notice that the different modalities of image, including image data of CT, MRI (magnetic resonance image), and PET (positron emission tomography), provide us various approaches to reprocess and extract more information data and finally make the utmost utility of them. Each modality has its own advantage, for example, on pulmonary nodules, different modalities were studied and CT were found to be most effective [24,25], other modalities still have their advantage, for example, MRI is non-radioactive and performed with multiple sequences which reveals more functional details of the lesion, it was limited due to motion artifacts and signals losses concerning respiratory motion and low proton density of lung but now become an important auxiliary assessment of pulmonary nodules [26]. We believe that in future radiomics used in different modalities may develop a series of comprehensive and sound methods that incorporate advantages of them and overcome the shortcomings.

Also we noticed that radiomics researches are no longer retrospective studies, but also extended its range to animal experimental studies [27], the positive result provide a strong evidence of the great potential in radiomics.

Potential and feasibility of radiomics on diagnosis of solitary pulmonary nodules

The application of radiomics for differentiating solitary pulmonary nodules is to find a series of characteristic features, which will show the significant difference between the malignant and the benign. We raise the hypothesis that radiomics are capable for differentiating the probability of malignancy of solitary pulmonary nodules by following reasons: first and foremost, malignancy and benign nodules are totally different neither in cell morphous nor biology behaviour. Since radiomics have been considered to have the potential on distinguishing different phenotypes of cancer. Differences between malignant nodules and benign ones are supposed to be revealed by radiomics features. Second, the natural high contrast between pulmonary nodules and lung parenchyma makes it an advantage in lung nodule segmentation. Many powerful segmentation algorithms are proposed, some of which were quite adequate [28-32]. An appropriate segmentation algorithm is the guarantee of satisfied reproducibility. Also an automatic or semiautomatic segmentation is a time saving helper, which makes it more usable if put into clinical practice. Third, more and more features have been proposed as well as various decision models which are developed on handling massive high-dimension data. These tools are design to interpret data that cannot be comprehended by human in quantity or dimension.

Studies of radiomics for differentiating pulmonary nodules have been performed and the extracted parameters such as texture features on CT scan were proved useful in differentiating transient from persistent part-solid nodules [33]. More and more features are being studied and finally being confirmed to be significant in different disease and playing different role in clinical practice [34]. These results have aggrandized the utilization of radiology and give a new sight of potential of radiomics. Most studies were performed on CT, and images were from different CT machines in some studies, which suggested the potential ability of radiomics to integrate information across machines. However, few studies on MR or PET showing positive results and these indicated the powerful information process abilities of radiomics across modalities [26,35]. MR are known to have the advantage of not only anatomic but also functional image compared with CT, and the reprocess of functional images may offer clinicians more useful information about metabolism, which would better the complementary job to raise the differential accuracy.

There is now no gold standard definition of pulmonary nodules available on MRI or PET. This mainly because that CT is recognized as the gold standard in detection of pulmonary nodules, although plenty of studies on detection the tissue with MRI or PET are also performed, most lung nodules are first detected with CT imaging, especially in recent popular lung screen programs, these programs using CT as a detecting instrument have led to an increasing amount of image data and thus confront us a challenge but also a chance [36]. To make the best use of these data thus becomes an urgent affair. Highly accurate diagnosis can both improve the prognosis of early lung cancer and reduce the unnecessary cost if the nodule does not need to resect. And for nodules that are not suitable for resection, monitoring the progression of the lesions after treatment then becomes the first to consider. Previous studies concerning lung nodules with diagnosis have been performed regarding to the detection, differentiation and classification of the disease. Detection, differentiation and classification are recognized as three levels in lung nodule assessmentand we use levels here rather than kinds because levels have the features that higher ones include lower one while kinds are distinct from each other. Differentiation is based on detection and also the foundation of classification. Differentiation will make no sense if detection rates stay relatively low, so does classification. We noticed that classification with radiomics in lung nodules shared a much larger fraction in recent studies, differentiating does not seem to be so favourable among researchers. We hereby emphasise the importance of differentiation, on which accurate classification should be based, and higher detection rates of lung nodules also call for better differentiation methods. Differentiating nodules between malignant and benign ones is an important issue since with earlier intervention of the malignant pulmonary nodules, they will get a better prognosis or even be cured, and it can also supply us with valuable information to the final protocol so as to confirm the different treatment decisions on lung nodules. Traditional radiology had done a lot on them (Table 1), but still there are some limitations [37]. For instance, we used to differentiating lung nodules by features such as morphology, size, density/intensity, some of which were evaluated by radiologists manually and subjectively. These visual features contain subjective information, which will inevitably introduce bias. Also, there is no standard for features such as morphology and size. Again, single or small numbers of descriptors are not likely to draw a picture of lesions, since the complexity in the development and progression of disease. Another question is how we can process these features and make them an accurate predictor. To date, scholars used different statistic models based on various features to predict the malignancies of pulmonary nodules [38]. Features including clinical data, location, size (e.g. diameter, doubling time, volume), computer CT densities, enhancement, significant signs (e.g. halo sign) and functional features (e.g. FDG/PET SUV, MRI ADC) were used as variances in these models; these models serve a great value [39-41]. The integration and the second use of these image data can maximize the usage of image information and finally allow us to get features fit each kind of disease and so a feature that can represent the potential biology behaviours is better. We are certain that some features can do this job, for example, to predict the distant metastasis in lung cancer, which represents the biology behaviour of the lesion, radiomics is an inevitable consequence and an adaption to the emergence of more detailed accuracy classification and more individualized treatments (Table 1).

Conclusion

Radiomics refers to a combination of radiology, bioinformatics, biomedical engineering, using imaging features as elemental units and utilize different algorithms and models as tools to analyse and transform these elemental units into useful, readable and reliable information. It appears with the advent of molecular medicine, individual therapy, which are core elements of the precision medicine. Its emergence brings a new way in understanding the etiology, pathology and progression of disease. More accurate and detailed information extracting from images, correlating with disease characteristics, are widely applied in diagnosis, surveillance, and treatment planning of diseases. Lung nodules are a common problem in clinical practice, differentiating malignant from benign is a substantial issue we need to address. Radiomics features with its high relation with cellular and molecular characteristics of lesions, are proved to be useful in differentiating pulmonary nodules, yet a number of features has been discovered significant in differentiating malignancy of lung nodules [13], and the number of robust features are increasing due to scientists’ hard work, these highly stable informative features are convinced to be closely correlated with the development and progression of disease, thus can serve a better job both in accuracy and specificity in differentiating pulmonary nodules. However, problems still exist in the way, such as how we integrate medical data and analyse it. Firstly, different machines or different protocols among hospitals may lead to the bias between images, in order to solve these problem, globally approved standard on scanning protocol has to be raised, which aimed at minimize the intra- and inter-device variety and thus improve the reproducibility and stability. Secondly, as radiomics studies contain so large amount of data, it is born with inevitable dependence on various algorithms, statistic methods and tests, but the processing of data is often too simple as lack of expertise in statistics, for example, many studies were deemed to inappropriate apply tests of significance (P-value) [47]. Proper application of these can improve the credibility of the results and serve a better performance as they are to be finally validated in clinical practices. As more and more new better algorithms and statistical methods are reported, we are able to build more accurate models in disease diagnosis and prediction. Refining the large amount of information is another problem since the large amount of data volume contains lots of both useful and useless information, to refine this information can help us to make a better usage of image data. Thirdly, although features extracted from images were found to be relative with molecular phenotype or clinical outcomes, how to illustrate these relationship leaves a huge room for us to investigate, some experts have tried this [45], and some features are explain to be with clinical signature, but still large number remains unknown, however, it still needs to be investigated in the long term with efforts from both clinical experts and mathematics professionals.

Future Perspectives

The advent of radiomics is more than an emerge of an available tools, but a potential new pattern in medicine, that is medical informatics, more and more advanced computer processing technology has made our life style changed a lot, so will it happen in medical procedure and every doctors mind, for example, those experience we are not able to share but only can be accumulate are now able to quantified. Proper algorithm and computer with powerful processing ability allow us to make more accurate and effective diagnosis and treatment plan than before.

Deciphering gene expression successfully with radiomics analysis is encouraging; however, it is just one of many functions of the radiomics. The potential application of radiomics is feasible and invaluable. Firstly, radiology are widely employed in clinical practice because of its non-invasion and acceptable costs, and now nearly every hospital is able to provide patients with radiologic examines. They are concerned as one of the initial and essential assessments gradually replacing some physical exams required after admission. Hundreds of thousands of image data is being created every second in every department of radiology all over the world, which supports us with a great number of data that contains valuable source information. In addition, radiology, including X-ray, Ultrasound, CT, MRI and PET are non-invasive, while most current cell-based pathological diagnosis methods are unavoidably invasive surgery or biopsy. Radiology provides us the chance to get information from patients without concerning pain and complications that are used to happen in invasive approaches, allowing us to collect medical image information repeatedly within a certain time. Besides, revolution of conceptual frame in data analysis, such as data-mining method, permits us to decipher and analyse series of high-throughput data avoiding the limitations of these techniques. Finally, radiology provides a unique approach and insight that render us to judge the lesion as a whole, to view, and to interpret biological behaviour or characteristics of the body tissues or organs. For example, heterogeneity caused by a large amount of genetic diversity within a tumor or among tumours is thought to play a weighty role in cancer growing pathway and results in phenotypic variations, which brings further challenges to cancer individualized therapy. It is demonstrated that heterogeneity, which reveals biology characteristics, drug-response and disease prognosis, is correlated with medical images by radiomics analysis [48]. The development of experimental medical science along with radiomics will give us the opportunity to interpret more information from imaging and thus improve the use of medical imaging and make it a more powerful tool that can give guidance to the whole clinical procedure whether in diagnosis, surveillance or prognosis.