Radiogenomics Based on PET Imaging

Abstract

Radiogenomics or imaging genomics is a novel omics strategy of associating imaging data with genetic information, which has the potential to advance personalized medicine. Imaging features extracted from PET or PET/CT enable assessment of in vivo functional and physiological activity and provide comprehensive tumor information non-invasively. However, PET features are considered secondary to features on conventional imaging, and there has not yet been a review of the radiogenomic approach using PET features. This review article summarizes the current state of PET-based radiogenomic research for cancer, which discusses some of its limitations and directions for future study.

Introduction

Radiogenomics, also called imaging genomics, is a rapidly evolving field of imaging science that seeks to uncover associations between image phenotypes and genomic information [1–4]. Image phenotypes or signatures are obtained by a radiomic approach, which is the extraction of qualitative or quantitative image features with or without computer algorithms [5]. Genomic information is much easier to use in clinical practice than 10 years ago, because of development about the advancements in high-throughput DNA sequencing technologies [6].

Radiogenomics is the cross-disciplinary research bridging imaging and genomics. The term radiogenomics has been used in different fields of research to refer to investigations of the underlying genetic causes of individual differences on radiation response, also called radiation genomics [7, 8]. We excluded radiation genomic studies and studies that examined the associations between imaging features and patient response to radiation therapy in this review.

In the era of personalized medicine that has focused on molecular characterization of cancer in individual patients, merging radiomic data and genomic data has great potential [9, 10]. Some genetic tests have been integrated in clinical practice of oncology since information about genetic alterations can enable physicians to select appropriate management strategies [11]. Tumor diagnosis is changing from the traditional histology-based approaches to gene-based stratification. However, routine application of genomic-based cancer characterization for personalized treatment is limited. Incorporating genomics into routine clinical practice is still a challenge because of the need for tissue samples collected via invasive procedures, high cost, relatively long turnaround time, and technical complexity [2, 12]. Even more importantly, the genetic information obtained from samples collected through biopsy or surgery may not accurately represent the whole genomic characteristics of lesions. Tumors have genetic diversity and heterogeneity not only among different masses but also within a single mass [11–14]. Intra- and inter-tumoral genetic heterogeneity eventually drives treatment failure and disease progression [14–16]. In addition, it is not feasible to perform multiple and sequential biopsies to identify genetic alterations within a single tumor lesion throughout the disease course.

Although histological classification of biopsied or surgically resected samples is the gold standard for oncologic diagnosis, in vivo tumor characterization by imaging is advantageous over invasive methods. Imaging techniques are non-invasive and can afford a more overall view of the entire tumor than that of biopsy samples alone. Furthermore, it can be performed repeatedly to evaluate treatment response and to monitor tumor progression in routine clinical practice [2]. The association between imaging features and gene expression enables valuable insight into which imaging features can be used to predict gene expression in tumors [2]. Associated features could be incorporated into multiparametric models for predicting mutations in particular genes or genetic pathways [17]. The radiogenomic approach primarily aims to identify imaging biomarkers for non-invasive genotyping. Bridging imaging and genomics may be useful in unraveling unknown molecular mechanisms underlying cancer progression.

Medical scans such as computed tomography (CT), magnetic resonance imaging (MRI), and positron emission tomography (PET) (or PET/CT) can be converted into quantitative data and incorporated into radiogenomic research. To date, radiogenomic approaches have focused on CT or MRI scans. However, PET or PET/CT has the advantage of allowing assessment of in vivo functional and physiological activity with higher accuracy than anatomical imaging modalities. Associations between image features extracted from ¹⁸F-fluorodeoxyglucose (FDG) PET/CT and genetic alterations have been investigated [18–22]. PET-based radiomic features have potential in predicting genetic alterations and could complement the radiogenomic approach using anatomical imaging modalities and improve the capability of multiparametric prediction models [9]. Although there have been many reviews of radiogenomic articles, PET-based radiogenomic research has only been briefly addressed as ancillary to CT- or MRI-based studies.

In this review, we discuss radiogenomic research based on PET imaging from the perspective of nuclear medicine. The purpose of this review is to provide a comprehensive view of the current state and further directions of PET-based radiogenomics. We assessed oncology-wide radiogenomic associations in previous studies, reviewed limitations of the radiogenomic approach, and searched for feasible solutions to these limitations.

Current State of PET-Based Radiogenomic Research

Study Search

We searched the electronic databases for original research published between 2000 and October 2019 that investigated the associations between PET imaging features and genomic data by entering a specific combination of search terms and conditions in MEDLINE. Details of the search strategy and process are provided in Fig. 1. Briefly, search terms included variations of different nuclear medicine imaging modalities such as “PET” and “tomography,” oncologic terms such as “tumor,” “cancer,” “carcinoma,” and radiogenomic terms such as “genetics,” “genomics,” “feature,” “index,” “radiogeno,” “radiomi,” and “imaging genomics.” We only included original studies with human subjects that were written in English. In addition, we identified reviews that contained the word radiogenomic or imaging genomics separately and reviewed all references to identify PET-based radiogenomic studies that were not included in our initial search. We excluded studies that correlated imaging features with biomarkers measured by immunohistochemistry or fluorescent in situ hybridization. For example, we did not include studies that correlated image features with molecular subtypes in patients with breast cancer.

Fig. 1.

Details of the search strategy and process

Finally, we included a total of 30 studies for this review (Fig. 1). Among them, there were studies that associated image features with a few specific genes. Given that omics are aimed primarily at the collective characterization and quantification of biological molecules [23], an association of one or several genes with image features was insufficient to accept as “real radiogenomics.” Therefore, similar to a previous review [2], we grouped them under the category “semi-radiogenomic studies.” As a result, 7 studies were categorized as radiogenomic (Table 1) and 23 studies as semi-radiogenomic (Table 2) studies.

Table 1.

Published PET-based radiogenomic studies (n = 7)

Authors	Year	Cancer	Study design	Data source	Number of subjects	Number of PET features	Feature type	Genome	Genomic analysis	Individual gene vs gene cluster	Pathway analysis	Validation dataset	Imaging modality
Osborne et al. [43]	2010	Breast	Pro	Single	20	1	I	RNA	Microarray	Both	Yes	No	PET/CT
Palaskas et al. [44]	2011	Breast	N/A	Single	18	1	I	RNA	Microarray	Cluster	Yes	Yes	PET/CT
Gevaert et al. [20]	2012	Lung (NSCLC)	Retro	Single	26	1	I	RNA	Microarray	Cluster	Yes	Yes	PET/CT, CT
Nair et al. [19]	2012	Lung (NSCLC)	Retro	Multi	25	14	I, H, V	RNA	Microarray	Both	Yes	No	PET/CT
Yamamoto et al. [22]	2016	Lung (NSCLC)	Retro	Public*	26	1	I	RNA	PCR	Individual	No	Yes†	PET/CT
Vlachavas et al. [24]	2019	Colorectal	Retro	Single	30	8	I, K	RNA	Microarray	Individual	No	No	PET
Crespo-Jara et al. [21]	2018	Metastatic	Retro	Single	84	7	I, V	RNA	Microarray	Both	Yes	Yes	PET/CT

*NSCLC Radiogenomics dataset that contained ¹⁸F-FDG PET and RNA expression profile dataset: Gene Expression Omnibus (GSE28827) & https://wiki.cancerimagingarchive.net/display/DOI/NSCLC+Radiogenomics%3A+Initial+Stanford+Study+of+26+Cases

†Internal validation

Table 2.

Published PET-based semi-radiogenomic studies (n = 23)

Authors	Year	Cancer	Study design	Data source	Number of subjects	Number of PET features	Feature type	Genome	Genomic analysis	Individual gene vs gene cluster	Pathway analysis	Validation dataset	Imaging modality
Kim et al. [33]	2010	Lung (NSCLC)	Pro	Single	154	1	I	DNA	SNP assay	Individual	Yes	No	PET/CT
Yoon et al. [32]	2015	Lung (ADC)	Retro	Single	539	1	I	DNA, RNA	Microarray	Individual	No	Yes*	CT, PET/CT
Guan et al. [29]	2016	Lung (NSCLC)	Retro	Single	85	1	I	DNA	PCR	Individual	No	No	PET/CT
Cho et al. [30]	2016	Lung (NSCLC)	Retro	Single	58	2	I, V	DNA	PCR	Individual	No	No	PET/CT
Yip et al. [28]	2017	Lung (NSCLC)	Retro	Multi	348	68	I, V, H, T	DNA	PCR, PROFILE Oncomap	Individual	No	No	PET/CT
Moon et al. [18]	2019	Lung	Retro	Single	176	106	I, V, H, T	DNA	NGS	N/A	No	No	PET/CT
Jiang et al. [27]	2019	Lung (NSCLC)	Retro	Single	80	256	I, V, H, T	DNA	PCR	Individual	No	No	PET/CT
Kawada et al. [36]	2012	Colorectal	Retro	Single	51	2	I, R	DNA	PCR	Individual	No	No	PET
Izuishi et al. [42]	2012	Colorectal	Retro	Single	37	1	I	RNA	PCR	Individual	No	No	PET
Miles et al. [38]	2014	Colorectal	Pros	Single	33	3	I	DNA	Pyrosequencing	Individual	No	No	PET/CT
Chen et al. [41]	2015	Colorectal	Retro	Single	103	4	I, V	DNA	PCR	Individual	No	No	PET/CT
Lee et al. [37]	2016	Colorectal	Retro	Single	179	4	I, V	DNA	PCR	Individual	No	No	PET/CT
Lovinfosse et al. [35]	2016	Colorectal	Retro	Single	151	20	I, H, T	DNA	PCR	Individual	No	No	PET/CT
Oner et al. [40]	2017	Colorectal	Retro	Single	55	4	I, V	DNA	PCR	Individual	No	No	PET/CT
Chen et al. [39]	2019	Colorectal	Retro	Single	74	65	I, H, V, T	DNA	PCR	Individual	No	No	PET/CT
Choi et al. [48]	2017	Thyroid (PTC)	Retro	Single	106	1	I	DNA	PCR	Individual	No	No	PET/CT
Chang et al. [49]	2018	Thyroid (PTC)	Retro	Single	108	1	I	DNA	Pyrosequencing	Individual	No	No	PET/CT
Nagarajah et al. [47]	2015	Thyroid (PTC)	Retro	Single	82	1	I	NA	NGS/ Sequenom	Individual	No	No	PET/CT
Metellus et al. [53]	2011	Brain (Glioma)	Retro	Single	33	1	R	DNA	PCR	Individual	No	No	PET/CT
Lanic et al. [54]	2012	DLBCL	Retro	Single	45	1	I	RNA	PCR	Cluster	No	No	PET
Kesch et al. [55]	2018	Prostate	Pro	Single	5	1	I		GWMA	CNA	No	No	MRI, PSMA PET/CT
Magometschnigg et al. [45]	2019	Breast	Retro	Single	67	1	I	DNA	Pyrosequencing	Individual	No	No	PET/CT
Liu et al. [51]	2016	Neuroblastoma	Retro	Single	42	2	I	RNA	PCR	Individual	No	No	FDG, FDOPA PET/CT

*Internal validation

I, intensity feature; R, uptake ratio; H, SUV histogram feature; V, volumetric feature; K, kinetic feature; NGS, next-generation sequencing; GWMA, genome-wide methylation analysis; CNA, chromosomal copy number alterations

Study Designs and Characteristics

Five of the seven radiogenomic studies (71.4%) were retrospective. One study used multi-institutional data, another study used public data, and the other five studies were conducted at a single institution. The number of patients ranged from 18 to 84, with a median of 26 patients. Five (5/7, 71.4%) studies adopted a validation dataset for verifying the association between image features and genetic alteration identified in the initial dataset. Among them, one study used an internal validation dataset. Different types of cancers were studied, i.e., lung cancer (3/7, 42.8%), breast cancer (2/7, 28.5%), colorectal cancer (CRC) (1/7, 14.3%), and metastatic tumors (18 different types) (1/7, 14.3%). The studies used different types of imaging modalities including FDG PET/CT (5/7, 71.4%), FDG PET (1/7, 14.3%), and FDG PET/CT and CT (1/7, 14.3%).

Twenty of the twenty-three semi-radiogenomic studies (86.9%) were retrospective. One study used multi-institutional data, and the remaining 22 studies were conducted at a single institution. The number of patients ranged from 5 to 539, with a median of 80 patients. One (1/23, 4.3%) study used validation data. Eight types of cancers were studied, i.e., CRC (8/23, 34.8%), lung cancer (7/23, 30.4%), papillary thyroid cancer (PTC) (3/23, 13.0%), breast cancer (1/23, 4.3%), glioma (1/23, 4.3%), diffuse large B cell lymphoma (DLBCL) (1/23, 4.3%), prostate cancer (1/23, 4.3%), and neuroblastoma (1/23, 4.3%). The imaging modalities used included FDG PET/CT (16/23, 69.5%), FDG PET (3/23, 13.0%), FDG PET/CT and CT (1/23, 4.3%), FDG PET/CT and 3,4-dihydroxy-6-(18)F-fluoro-l-phenylalanine (FDOPA) PET/CT (1/23, 4.3%), and MRI and gallium 68-labeled prostate-specific membrane antigen (Ga-PSMA) PET/CT (1/23, 4.3%).

Imaging Features

Seventeen of the thirty studies (56.6%) performed PET image acquisition with one scanner, and twelve (12/30, 40.0%) studies used two or more scanners. One study did not provide any information about the scanner. Sixteen (16/30, 53.3%) studies extracted PET image features from volume of interest (VOI), and 7 (7/30, 23.3%) studies used region of interest (ROI). Seven studies did not provide detailed information regarding target lesion segmentation. VOI or ROI were automatically or semi-automatically drawn in 14 studies (14/30, 46.6%), and they were manually drawn in 9 studies (9/30, 30.0%). The number of imaging features extracted ranged from 1 to 256 with a median of 2. Imaging features included intensity features such as SUV_max, volumetric features such as MTV and TLG, SUV histogram–based features, uptake ratio, texture features, and kinetic parameters extracted for compartment model (one study [24] performed dynamic PET study). SUV_max was the most common PET image feature extracted (29 studies (29/30, 96.6%)). Sixteen studies (16/30, 53.3%) used only intensity features, five studies (5/30, 16.7%) used intensity and volumetric features, and five studies (5/30, 16.7%) used intensity, volumetric, histogram, and texture features. The PET image features used in the remaining five studies were intensity and uptake ratio (n = 1, 3.3%); intensity and kinetic features (n = 1, 3.3%); uptake ratio (n = 1, 3.3%); intensity, SUV histogram, and volumetric features (n = 1, 3.3%).

Genomic Data

Seventeen (17/30, 56.6%) studies extracted data from DNA, 10 studies (10/30, 33.3%) extracted data from RNA, 1 study (1/30, 3.3%) used data from RNA and DNA, and 1 study (1/30, 3.3%) used genome-wide methylation analysis data. One study did not provide detailed information about the genomic data source. Eleven of the seventeen studies that used DNA used polymerase chain reaction (PCR) analysis, three studies used pyrosequencing, one study used next-generation sequencing (NGS), one study used single nucleotide polymorphisms (SNP) assay, and one study used mass spectrometry genotyping and PCR. Among the ten studies that used RNA, six used microarray and four used PCR. The study with DNA and RNA used microarray analysis.

Three (3/30, 10.0%) studies classified individual gene expression into gene clusters to associate with image features; 22 (22/30, 73.3%) studies directly associated individual genes with imaging features; and 3 (3/30, 10.0%) studies used both approaches. One study used genetic heterogeneity and mutation burden index, and one study used chromosomal copy number alteration. Six (6/30, 20.0%) studies used pathway analysis in the initial gene clustering or in the final analysis for identifying significant genomic markers.

Lung Cancer

Lung cancer is the most actively studied in PET-based radiogenomic research. Ten of the thirty total studies included in this review focused on lung cancer. Three of these studies with microarray data were classified as “radiogenomic” research, and the remaining studies were classified as “semi-radiogenomic” research.

The three “radiogenomic” studies used a similar research approach. Gevaert et al. demonstrated a radiogenomic strategy to identify prognostic imaging biomarkers by leveraging public gene expression microarray data [20]. They extracted CT image features and SUV_max from CT and FDG PET/CT, and made a radiogenomic map by correlating the imaging features to metagenes, which are aggregated patterns of gene expression. By using the radiogenomic correlation map as an intermediate, they linked imaging features expressed in terms of metagenes to the public gene expression dataset where survival outcomes are available [20]. Their approach was able to leverage imperfect datasets, imaging, and a gene expression dataset without survival outcomes to draw new conclusions with imaging biomarkers for prognosis [20]. Nair et al. used a similar strategy linking a radiogenomic association map to public gene expression data with survival outcomes. They demonstrated that four genes (LY6E, RNF149, MCM6, FAP) associated with FDG uptake were also associated with survival. They performed SAM analysis to define genes significantly associated with FDG uptake metrics. The array data was filtered based on a significant detection call in at least 60% of the samples and log transformed using quantile normalization to account for array variation [19]. Notably, the prognostic metagene signatures associated with multivariate FDG uptake features were highly associated with survival outcome in the external and validation cohorts [19]. Yamamoto et al. demonstrated that non-small cell lung cancer (NSCLC) with high normalized SUV_max may be associated with epithelial-mesenchymal transition, which is implicated in metastasis and chemoresistance [22]. They used an integrative radiogenomic approach by linking to an external public dataset, similar to the strategy described above.

The research focus in the semi-radiogenomic studies was on identifying correlations between image features and clinically relevant mutations. The most researched genetic mutation in NSCLC was epidermal growth factor receptor (EGFR) and most commonly used methods for detecting EGFR mutations was PCR. The correlation of imaging features with EGFR is of interest because the genetic mutation status of EGFR plays a crucial role in treatment selection for NSCLC [25]. Receptor blockers or tyrosine kinase inhibitors (TKIs) are treatment options in patients with NSCLC with an activating EGFR mutation [26]. Previous studies demonstrated that PET features were significantly associated with EGFR mutation status [27–30]. A model using 35 selected features significantly predicted EGFR mutation status (area under the ROC curve = 0.953) [27]. Yip et al. demonstrated that PET-based radiomic features were significantly associated with EGFR mutation status, while none of the features was associated with KRAS mutation status [28]. Yoon et al. investigated the predictive value of PET features for anaplastic lymphoma kinase (ALK) rearrangements, which is one of the major therapeutic targets in a molecularly defined subset of NSCLC [31]. ALK fusion–positive tumors had significantly different SUV_max than c-ros oncogene 1 (ROS)/rearranged during transfection (RET) fusion–positive tumors [32]. The polymorphisms in the glucose transporter gene were associated with SUV_max in combination with hypoxia-related genes in patients with squamous cell carcinoma [33]. One study investigated the correlations between PET radiomics and genetic properties such as heterogeneity and tumor mutation burden [18], and revealed that surface SUV_entropy is significantly correlated with genetic heterogeneity in small cell lung cancer.

Colorectal Cancer

Nine searched studies focused on CRC. One study was “radiogenomic” research and the remaining eight studies were “semi-radiogenomic” research. In the radiogenomic study that used microarray data and PET data feature sets, Spearman’s correlation analysis was used to select the most correlated genes, from the initial list of the 911 common genes, with any of the PET variables. A composite signature was derived using seven genes and five PET variables, and evaluated for its discrimination performance between primary CRC and adjacent normal mucosa [24]. This radiogenomic signature had superior classification accuracy [24].

As in lung cancer, semi-radiogenomic studies of CRC tested PET imaging features to predict clinically relevant gene mutations. Anti-EGFR therapy is an important treatment option for advanced CRC, and an activating Kirsten rat sarcoma viral oncogene homolog (KRAS) mutation is considered indicative of resistance to anti-EGFR therapy [34]. Therefore, the presence of KRAS mutation serves as a predictive biomarker in some CRC patients [34]. In the semi-radiogenomic studies, the most commonly used methods for detecting KRAS mutations was PCR. Although PET-based radiomic markers have been tested for predicting KRAS mutation, the results were inconsistent. SUV intensity features such as SUV_max [35–38], SUV histogram features [35], and volumetric features [39] were significantly associated with KRAS mutation in some studies. On the other hand, other studies reported that SUV_max is not indicative of KRAS mutational status [40, 41]. SUV_max was also tested in correlation analysis with expression of other genes [42], and was significantly correlated with hypoxia-inducible factor alpha (HIF1α) expression.

Breast Cancer

Among the three studies of breast cancer included in this review, two used genome-wide microarray data and one used individual gene mutation data. Molecular subtyping of breast cancer based on expression of estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor markers (HER2) has direct prognostic and therapeutic implications in patient management [43]. As in lung cancer and CRC, studies in breast cancer focused on identifying the correlation between PET image features and status of clinically relevant mutations. Osborne et al. found an association between ER status and increased glucose uptake presented as SUV [43]. ER-negative tumors had significantly higher SUV_max than ER-positive tumors. However, there was no significant association between SUV and PR status or HER2 status. Furthermore, they established a preliminary radiogenomic gene set whose expression mostly correlated with ER status and SUV. They screened microarray data to find genes associated with increased SUV and compared these genes with those of a previously validated ER status molecular phenotype to identify significantly overlapping genes. Palaskas et al. identified genetic pathways associated with SUV in breast cancer, which were the coordinately upregulated multiple metabolic pathways [44]. They showed that these pathways could serve as an “FDG signature” which was closely linked to the “basal-like” intrinsic breast cancer subtype and activation of the MYC oncogene [44]. A recent study of individual mutational analysis investigated the association of PIK3CA mutational status, a novel therapeutic target, with SUV_max in invasive breast cancer [45]. No significant correlation between SUV_max and PIK3CA gene mutation was found in that study.

Other Malignancies

In PTC, the BRAF V600E gene mutation is an important prognostic factor for more rapid cancer growth and a higher death rate [46]. An association between SUV_max and BRAF V600E mutation has been demonstrated with consistent results [47–49]. MYCN amplification is a well-known genetic marker for poor prognosis in patients with neuroblastoma [50]. Liu et al. investigated the utility of FDG PET/CT and FDOPA PET/CT in determining the prognosis of neuroblastoma and reported that SUV_max of FDG and FDOPA PET had a significant association with MYCN amplification [51]. Gliomas with mutated IDH1 and IDH2 have better prognosis than those with wild-type IDH [52]. IDH mutation status has no significant association with FDG PET features in World Health Organization (WHO) II and III gliomas [53]. Lanic et al. assessed the relationship between molecular subclassification of diffuse large B cell lymphoma (DLBCL) by gene expression profile and PET features [54]. The expression of genes encoding glucose transporters from SLC2A1 to SLC2A5 was assessed and correlated with baseline SUV_max. Among these genes, SLC2A3 appeared to be the most highly expressed [54]. Kesch et al. assessed chromosomal copy number alterations (CNAs), a genetic marker of tumor aggressiveness, and MRI and ⁶⁸Ga-PSMA-11-PET/CT features in prostate cancer [55]. Radiogenomic analysis was performed and CNAs were directly correlated to imaging features. There was a strong relationship between the radiologic and genomic index lesions, which suggests that imaging features can be used to identify the genomically most aggressive region within the prostate [55]. Crespo-Jara et al. built the universal genomic signature which could predict the increased FDG uptake in various metastatic tumors through a radiogenomic approach [21].

Limitations and Directions for Future Research

Study Design, Sample Size, and Validation

There are few studies that used large-scale genome data. Seven of the thirty studies that we identified (7/30, 23.3%) were classified as real radiogenomic studies in the sense that they used genome-wide gene expression data. Most of the remaining studies tested the association between one or several clinically known genes and PET imaging markers, and we classified these as semi-radiogenomic studies. The semi-radiogenomic approach may be an efficient method to identify clinically relevant imaging-to-genomics relationships. However, it may undermine the unknown potential radiogenomic signatures. Given that the primary goal of radiogenomic analyses is to uncover the association between image phenotypes and genomic information, it would be advantageous to include as many indicators as possible. Nonetheless, radiogenomic research with whole genome data has the limitation of dimensionality [25], which arises when analyzing high-dimensional data, often with more than hundreds of dimensions, that does not occur in low-dimensional settings [56]. As the number of data variables grows, the amount of data needed to generalize accurately grows exponentially [25, 56, 57]. Since in reality it is almost impossible to collect this many samples, real radiogenomic research needs to reduce the dimensionality of genomic data. A common solution for the dimensionality problem is to group individual genes into clusters or metagenes before performing association analysis with imaging features [19, 20, 25]. The problem of dimensionality also occurs when analyzing imaging feature data. However, in general, the dimensionality of imaging data is lower than that of whole genome data. Previous radiogenomic studies with PET often used only a few imaging features, despite the many available features. The number of imaging features extracted ranged from 1 to 14 with a median of 1 in the radiogenomic studies (n = 7), which may further undermine the potential PET imaging-to-genomics association. On the other hand, extraction of fewer image features reduces dimensionality, which facilitates accurate data analysis.

Previous real radiogenomic studies are also limited by small sample size (range 18 to 84, median of 26 patients), which decreases the power of the study. Sample sizes that are too small increase the likelihood of a type II error skewing the results. Therefore, it is possible that some of the imaging-to-genomics relationship was not detected in previous studies. In addition, these studies also suffer from lack of a validation dataset verifying an identical statistical association. A validation dataset was used in four of the seven real radiogenomic studies (4/7, 57.1%) that we identified, and in only one of the twenty-three semi-radiogenomic studies (1/27, 3.7%). In fact, the difficulty in obtaining original cohorts of patients with both appropriate PET scans and adequate tissue samples for genomic analysis resulted in small sample sizes and lack of validation [2, 4]. Although large-scale imaging data are available, many of which have corresponding tumor tissue available for various molecular analyses [2], and genome-wide data are costly and technically complex [2, 12]. However, the cost of high-throughput molecular techniques including next-generation sequencing has been reduced to a fraction of what it was before [2, 58]. This technical advance has made it easier to obtain large amounts of genetic data. Genome imaging datasets can be collected retrospectively or prospectively in relatively large quantities. However, depending on the cancer type, PET/CT is not included in clinical practice at initial diagnosis where biopsies are routinely performed. Therefore, unlike CT, it is not practical to harvest large amounts of genome PET data to create a training set and validation set of sufficient size.

Use of publicly available resources is a feasible solution to overcome the scarcity of large datasets containing genomic data and PET images. The Cancer Genome Atlas (TCGA) is a rich dataset containing multidimensional genomic and clinical data for multiple types of adult cancers [59]. The Cancer Imaging Archive (TCIA) is another resource for image data accessible to the public and provides image files mainly in Digital Imaging and Communications in Medicine (DICOM) format. TCIA is a large archive of medical images that includes cases corresponding to the patients in the TCGA [60]. This database can serve as powerful validation tool and facilitate the discovery of underlying relationships between cancer genomics and medical image features. However, a validation study with data from external resource requires addressing data availability issues, and the utility of TCIA for PET-based radiogenomic research is limited because of the small number of PET images comprising only a fraction of the total image data contained in the TCIA. More importantly, it is also limited by the fact that gene expression profiles cannot be matched to a specific location on imaging in some cases [2, 60].

In addition to TCIA, researchers can take advantage of independent datasets produced by other research groups. For example, Bakr et al. developed a unique radiogenomic dataset from a cohort of 211 patients with NSCLC that comprises CT and PET/CT images, semantic annotations of the tumors, and segmentation maps of tumors. They also matched the imaging data with results of genetic analysis and other clinical information [61].

Vulnerabilities of PET Radiomic Features

PET imaging features are significantly affected by several factors [18, 62–66]. One of the key factors influencing PET imaging feature extraction is the size of the target tumor lesion [63]. PET texture features that represent heterogeneity of the target lesion are particularly vulnerable to the size. Brooks et al. reported that, for texture analysis of FDG PET, it is appropriate to have a tumor volume of at least 45 cm³ [64]. They showed that the features are five times more sensitive to volume changes at small volumes below the proposed minimum than in those above it [64]. However, the larger study demonstrated that minimum metabolic tumor volume for analysis can be as small as 10 cm³ [67]. Image acquisition parameters and the reconstruction method are also important factors affecting the extraction of PET image features. Although there was a difference in degree, extracted FDG PET features vary depending on these conditions [18, 62, 65]. Galavis et al. demonstrated that 40 of 50 features (80.0%) had large variation above 30% in a study where PET scans were acquired in both 2D and 3D mode, and reconstructed using five different reconstruction methods [65]. The tumor segmentation method also affects PET image features [18, 66]. For example, a previous study reported that the coefficient of variance of 70 of 106 features obtained using nine different segmentation methods was above 0.1 [18]. Consequently, unless PET features are obtained under similar study conditions, these kinds of radiogenomic studies are difficult to compare with each other [13]. The repeatability of feature extraction is also influenced by imaging acquisition conditions [68, 69]. This means that even PET features obtained under the same conditions may differ. The vulnerability of PET radiomic features is the major obstacle to generalization of previously reported genome-image relationships.

To overcome the lack of reproducibility in high-throughput image analysis studies, researchers have been working towards standardization of imaging feature extraction [70–72]. Due to the high susceptibility to various factors, the use of PET quantitative features as clinical imaging biomarker requires a highly standardized imaging acquisition and preprocessing protocol. Various factors affecting image features were identified, and recommendations to take these factors into account were proposed in the literature [70]. Most previously published recommendations seem to be consistent [70]. Existing recommendations and guidelines are incorporated into the reference manual of the image biomarker standardization initiative (IBSI), which is an independent international collaboration for the standardization of radiomics. The IBSI provides benchmark datasets and benchmark values to verify image processing and image biomarker calculations, as well as consensus-based guidelines including imaging feature nomenclature, definitions, and reporting system, for high-throughput image analysis of CT, MRI, and PET [71, 72].

Another possible way to overcome the variability of PET image features is to use stable features. Although almost all PET features are affected by various factors, there are the features that are robust to the choice of imaging settings and are insensitive to variation [68–70, 73, 74]. Shape features, area under the curve of the cumulative histogram, and histogram-derived entropy were very reliable first-order features (with less than 5.0% of standard deviation) in a repeatability analysis of a prospective multicenter cohort study. Entropy_{gray-level-cooccurrence-matrix features (GLCM)}, sum entropy_GLCM, and difference entropy_GLCM were the most repeatable second-order features. Small zone size emphasis and zone size percentage were reliable third-order features [75]. A small number of features were found to be repeatable (intraclass correlation coefficient = < 0.8) for size, FDG uptake, image reconstruction, noise, discretization method, and segmentation in a repeatability analysis with a phantom study. Intensity and entropy of second-order features were also repeatable [68]. In another study, the inverse difference moment GLCM was insensitive to variations in segmentation, discretization, and reconstruction methods [69]. Low gray level run emphasis (3rd-order feature), energy, maximal correlation coefficient (2nd-order feature), and entropy (1st-order feature) showed small variability (range ≤ 5%) according to different acquisition modes and reconstruction methods [65]. In addition, first-order features are less affected by tumor volume and provide reproducible values [73, 74]. PET imaging features with high reproducibility and low variability under different conditions are better candidates for radiogenomics. However, previous studies have some discrepancies about which features are robust, thus they must be carefully selected.

The other way for lowering the variability of PET radiomic features is to use a statistical method. PET images are affected by scanner, acquisition protocol, and reconstruction parameters, thus hindering multicenter studies. Orlhac et al. developed a post-reconstruction harmonization method based on Bayesian framework to adjust the multicenter effect for textural features and SUVs [76]. This method is likely to be efficient at removing the technical heterogeneity arising from different conditions of institutions where the research is conducted, which could facilitate the multicenter studies and external validation of PET radiomic models [76].

Implementation and maintenance of the effort to overcome the issues mentioned above will be increasingly important in this research field. It is likely that compliance with the methods will be assessed as a measure of research quality. For example, to evaluate the validity and completeness of radiomic studies, radiomics quality score (RQS) was proposed by Lambin et al. [77]. RQS consists of 16 key components such as robust segmentation, use of standardized imaging protocols, and use of validation dataset. Each component was assigned scores according to its importance [77].

Conclusion

Radiogenomics, linking of imaging phenotypes to distinct molecular phenotypes of cancer, can impact the treatment and prognosis of a wide range of human cancers and will help advance personalized medicine. In this review, we examined all the major radiogenomic studies using PET imaging and investigated the current status of this field of research. We discussed the limitations of PET-based radiogenomic studies and tried to address feasible solutions to these limitations. While PET imaging features have potential in radiogenomics, only a small number of radiogenomic studies have been published. The majority of studies were semi-radiogenomic, relating expression of specific genes to imaging features. Small study samples and lack of validation due to the scarcity of large datasets and vulnerabilities of PET radiomic features under different conditions are major challenges that still need to be addressed. However, with publicly available data sources, standardization of feature extraction, and careful selection of robust features, future studies could overcome many of the challenges impeding the advance of a PET-based radiogenomic approach.

Funding Information

This work was supported by a National Research Foundation of Korea (NRF) grant funded by the Korean government (MSIP) (No.NRF-2016R1C1B2013411).

Compliance with Ethical Standards

Conflict of Interest

Yong-Jin Park, Mu Heon Shin, and Seung Hwan Moon declare that they have no conflict of interest.

Ethical Approval

This article does not contain any studies with human participants performed by any of the authors.

Informed Consent

Not applicable