Encyclopedia
Colorectal carcinoma is characterized by intratumoral heterogeneity that can be assessed by radiogenomics. Radiomics, high-throughput quantitative data extracted from medical imaging, combined with molecular analysis, through genomic and transcriptomic data, is expected to lead to significant advances in personalized medicine.
- radiogenomics
- machine learning
- deep learning
- oncologic imaging
- genomics
1. Introduction
Worldwide, with over 1.8 million new colorectal cancer cases and 881,000 deaths, colorectal cancer ranks third in terms of incidence (10.2%) and second in terms of mortality (9.2%) [1]. Colorectal cancer (CRC) is a heterogeneous disease in terms of etiology, biology, therapy response, and prognosis. Environmental factors (e.g., high intake of red and processed meats, sugars, highly refined grains, smoking and heavy alcohol use, obesity), personal history of cancer, chronic inflammatory bowel diseases and age are well known examples of factors related to the increased risk of CRC [2]. Colorectal adenocarcinomas are the result of a stepwise progression from normal tissue epithelium to carcinoma. TNM classification (American Joint Committee on Cancer and the Union for International Cancer Control; TNM classification (T-Tumor; N-Node; M-Metastais), the most used classification for CRC, helps stratify patients into different therapeutic and prognostic subgroups), the most used classification for CRC, helps stratify patients into different therapeutic and prognostic subgroups [3]. The 5-year relative survival rate for CRC ranges from 90% in patients diagnosed with stage 1 disease to 70% for patients with regional spread, and down to 10–12% for patients with stage IV disease [4,5]. Within the same TNM stage, survival rates remain highly variable for different patients, which can be explained in part by tumor biological heterogeneity. Molecular biology showed that multiple events involving tumor-suppressor genes, oncogenes, and DNA mismatch repair genes contribute to the development of CRC, and the order in which these mutations occur is critical [6]. This results in spatial tumor heterogeneity (different tumoral clones within the primary tumor), intertumoral heterogeneity (tumoral heterogeneity between patients with the same histologic type), and temporal heterogeneity(differences that developed inside the tumor over time) [7].
Recently, tumor heterogeneity has also been assessed by radiomic features, i.e., quantitative metrics extracted from medical images. Different studies have investigated the prognostic value of radiomic features of patients with colorectal cancers derived from different modalities (e.g., PET, CT and MRI). Radiogenomics investigate the relationship between imaging features and gene expression alterations, and/or their potential added or complementary value in predictive oncological modeling.
2. Radiomics
2.1. Radiomics Workflow
Medical imaging plays a central role in the detection, diagnosis, staging, and treatment of cancer and can provide morphological, structural, metabolic, and functional information [8]. The use of robust machine-learning techniques allows the extraction and exploitation of high-dimensional mineable data (morphologic, intensity-based, fractal-based, and textural features) from medical images for a more exhaustive characterization of tumor phenotypes [9]. The complexity of the radiomics workflow increases the need for standardized nomenclature and computation methods, which is being addressed by the Image Biomarker Standardization Initiative (IBSI). The IBSI currently provides standardized image biomarker nomenclature and definitions, a standardized general image processing workflow, tools for verifying radiomics software implementations, and reporting guidelines for radiomic studies [10]. Radiomics quality score (RQS) offers also a tool to analyze the methodology, reproducibility, and clinical applicability of radiomics studies [11].
Different modalities of medical imaging (CT [12], MRI, PET) are utilized for radiomics and radiogenomics studies. Before feature extraction can be carried out, the definition of a region/volume of interest (ROI/VOI) must be realized. It is one of the most challenging tasks in medical image analysis, especially in colorectal cancer, due to the size and shape of the organ and lesions (in any modality), as well as limited contrast between tissues (mostly in CT). The segmentation techniques of medical images are often specifically optimized for each combination of application, imaging modality, and body part [13]. In CT, because of the small intensity differences between primary colorectal tumor and adjacent tissues, the majority of studies relied upon a manual or semi-automated segmentation [14]. There are advantages and disadvantages to each approach. Manual segmentation requires expert annotation and is tedious, time-consuming, and prone to inter- and intra-observer variability. Semi-automatic segmentation methods have higher repeatability but may not always be as accurate as the manual segmentation in some situations (e.g., delineating the rectal tumor after neoadjuvant therapy) [15]. Automated segmentation is also subject to (more limited) variability. Therefore, the robustness of the radiomic features issued from multiple segmentations should be assessed in the workflow [16]. Fully automated segmentation methods, supervised and unsupervised, are generally built on basic image processing of pixel intensities and/or textural features, with the most promising methods relying on deep learning by training a U-net type structure [17]. Supervised techniques are considered to be more accurate but interobserver variability will still be present, as the manual part of the segmentation and the settings of the algorithm influence the result [12,18]. Unsupervised segmentation techniques commonly rely on labeled atlases and have been shown to be less accurate than the supervised techniques [19].
2.2. Features Extraction
Quantitative imaging features are subsequently extracted from the previously identified VOI. The features are usually classified into a number of families, such as intensity histogram-based features, shape-based features, texture-based features [20]. Several features require additional image processing steps before feature calculation (e.g., some intensity histogram metrics and all textural features require a prior discretization of intensities into a determined number of bins) [10].
2.3. Radiomics in Colorectal Cancer
Statistical analysis and prognostic model building are the next steps in radiomics analysis. Because a large number of radiomic features can be extracted from the image datasets, most of these exhibit intercorrelation and therefore are redundant. Combined with a small sample (i.e., number of patients), this can often contribute to the over-fitting of the models [21]. Feature selection should be performed to identify the most relevant and non-redundant subset of features that will be exploited to train a multiparametric prediction model according to the clinical endpoint [22]. Advanced machine and deep learning algorithms are employed for training/validation and testing of radio(geno)mics models. Various machine learning algorithms can be utilized—multivariate regression with least absolute shrinkage and selection operator (LASSO). LASSO serves as a regularization and variable selection method for any statistical model by shrinking the regression coefficients, and reducing some of them to zero [23]. Random forests (RF) is a machine learning algorithm that combines the predictions of a large number of small decision trees to produce a more accurate prediction [24]. Support vector machine (SVM) is an algorithm that performs classification by finding the hyperplane that maximizes the margin separating the classes [25].
Different studies have investigated the prognostic role of radiomics analysis in CRC. Radiomic analysis of colorectal cancer showed correlations with lymph node metastasis [26] and perineural invasion [27] or prediction of lung metastasis [28]. For colorectal liver metastasis, radiomics showed promising performance in predicting overall survival [29], disease-free survival [30], response to targeted therapies [31], the outcome for patients with unresectable hepatic metastases [32], or recurrence after liver resection [33]. In rectal cancer, radiomics were shown to be correlated with tumor regression [34], lymph node metastasis [35,36] or response after chemoradiation therapy [37,38].
3. Genomics and Transcriptomics
The development of CRC, from adenoma to adenocarcinoma, is the result of accumulated mutations in multiple genes that regulate cell growth and differentiation [39]. Colorectal carcinogenesis can arise from one or a combination of three different conditions, namely chromosomal instability (CIN), CpG island methylator phenotype (CIMP), and microsatellite instability (MSI) [40]. The classical CIN, observed in 65–70% of sporadic colorectal cancers, is associated with the acquisition of mutations in the adenomatous polyposis coli (APC), mutation of the KRAS oncogene, loss of chromosome 18q and deletion of chromosome 17p, which contains the important tumor suppressor gene TP53 [41]. The consequence of CIN is an imbalance in chromosome number (aneuploidy), subchromosomal genomic amplifications, and a high frequency of loss of heterozygosity (LOH) [42]. CIN + CRCs are correlated with poorer survival irrespective of ethnic background, anatomical location, and treatment with 5-FU [43].
CpG island methylator phenotype CRC accounts for 15–20% of sporadic CRC and characterized by the vast hypermethylation of promoter CpG island sites, resulting in concomitant hypermethylation of multiple genes, silencing of normal tumor-suppressor function and cancer formation [44]. Many genes that have been identified to be affected in CIMP have important functions in the cell, (e.g., CACNA1G, IGF2, NEUROG1, SOCS1 and RUNX3) [45]. CIMP-high CRCs are associated with distinct clinicopathological and molecular features such as older age, female preponderance, proximal tumor location, higher grade, reduced COX-2 expression, increased frequency of TGFBR2 mutations, and high rate of MSI, KRAS, and BRAF mutations [46].
Microsatellite instability (MSI), detected in about 15% of all colorectal cancers, is a molecular phenotype due to a defective DNA mismatch repair system [47]. The MSI-high phenotype is characterized by mucinous or signet ring appearance, poor differentiation, proximal colon, prominent lymphocytic infiltration [48].
Transcriptomics is the study of the RNAs transcription and expression levels, functions, locations, trafficking, and degradation. RNAs studies allow the identification of genes that are differentially expressed in distinct cell populations or in response to different treatments [49]. The CRC Subtyping Consortium identified four molecular subtypes with distinguishing features: CMS1 (microsatellite instability immune, 14%), hypermutated, strong immune activation and microsatellite unstable; CMS2 (canonical, 37%), epithelial, marked WNT and MYC signaling activation; CMS3 (metabolic, 13%), epithelial and metabolic dysregulation; CMS4 (mesenchymal, 23%), prominent transforming growth factor–β activation, angiogenesis, and stromal invasion [50].
4. Radiogenomics in Colorectal Cancer
The molecular landscape of CRC has prognostic relevance and affects the choice of therapeutic strategies [51]. Carcinogenesis could be triggered by the activation of several pathways downstream the epidermal growth factor receptor (EGFR), (RAS/MAPK; SRC/FAK; PI3K/AKT pathway) through deregulation of protein synthesis, cell cycle, apoptosis, angiogenesis [52]. KRAS mutations are present in 30–50% of colorectal cancers and the RASCAL study of 2721 colorectal cancers showed that the presence of KRAS mutation was significantly associated with poorer prognosis [53]. K-Ras is a critical mediator of EGFR-induced signaling cascades and resistance to anti-EGFR therapies has been observed in patients with KRAS mutation [54]. Cetuximab and panitumumab, EGFR monoclonal antibody-based therapies that block ligand binding and lead to the inhibition of the downstream RAS-RAF-MEK-ERK signaling pathway, are reserved for patients with wild-type KRAS mCRC [55]. MSI-H non-metastatic CRC patients have improved survival and receive no benefit from fluorouracil (FU)-based adjuvant therapy [56]. Therapy with programmed cell death 1 (PD1)-blocking antibodies, pembrolizumab, and nivolumab have shown efficacy in patients with MSI metastatic CRC [57]. Other pathways implicated in initiation, progression, activation, and migration of CRC, such as Wnt/β-catenin, Notch, Hedgehog, and TGF-β/SMAD, (PI3K)/AKT could be potential sites for targeted therapy [58].
Since genomic analysis is now essential for therapy in colorectal cancer, there have been several attempts to explore a potential role of radiomics within this context, either as a surrogate of genomics (i.e., “virtual biopsy”) or as a complementary tool (i.e., added information). The development of radiogenomics models capable of predicting CRC genetic mutations is very useful in general practice to improve decision-making and patient outcomes.
This entry is adapted from the peer-reviewed paper 10.3390/cancers13050973
