Radiomics in esophageal and gastric cancer

Esophageal, esophago-gastric, and gastric cancers are major causes of cancer morbidity and cancer death. For patients with potentially resectable disease, multimodality treatment is recommended as it provides the best chance of survival. However, quality of life may be adversely affected by therapy, and with a wide variation in outcome despite multi-modality therapy, there is a clear need to improve patient stratification. Radiomic approaches provide an opportunity to improve tumor phenotyping. In this review we assess the evidence to date and discuss how these approaches could improve outcome in esophageal, esophago-gastric, and gastric cancer.

As an example, data from the OEO2 and MAGIC trials for esophageal and esophago-gastric cancer have shown a 6% [3] and 13% [4] improvement in 5-year overall-survival, respectively. Trimodality therapy for esophageal and esophago-gastric cancer combining neoadjuvant chemo-and radiation-therapy in addition to surgery may also be superior to neoadjuvant chemotherapy alone in a selected patient population [8]. The CROSS trial [6] comparing neoadjuvant chemoradiotherapy plus surgery with surgery alone in patients with esophageal and esophago-gastric cancer showed a superior overall-survival of 49 vs. 24 months, hazard ratio 0.657, p = 0.003, and a pathological complete response rate of 29%, for patients with multi-modality treatment with no increase in surgical mortality (4% in surgery and trimodality groups, respectively).
A recent systematic review and meta-analysis of neoadjuvant chemotherapy in patients with gastric cancer has also found improved 3-year survival rates (relative risk 1.30; 95% CI 1.06-1.59, p < 0.01) [9]. Typical management pathways are shown in Figs. 1 and 2 for esophageal, esophago-gastric, and gastric cancer, respectively. Nevertheless overall-survival remains poor despite these improvements in patient care.
Recent genomic analyses have highlighted the genetic heterogeneity present in esophageal, esophago-gastric [10], and gastric cancer [11,12] as an underlying cause for the differences in outcome and heterogeneity of response to therapy. Quality of life also remains poor for many patients post-surgery, taking up to 3 years to return to pre-therapy levels in patients undergoing esophageal resection [13]. Better patient stratification remains a key challenge for patients with upper gastrointestinal tract cancers.

The imaging pathway at staging
For esophageal and esophago-gastric cancer, contrast-enhanced computed tomography (CT) remains the most commonly performed first step in staging due to the high prevalence of metastatic disease at presentation [14][15][16]. For patients being considered for a curative pathway, endoscopic ultrasound (EUS) and 18F-fluorodeoxyglucose positron emission tomography/CT (18F-FDG PET/CT) are performed due to the high sensitivity and specificity of EUS for local tumor and nodal staging; and 18F-FDG PET for distant metastases [17][18][19]. This aims to reduce the futile surgery rate.
For gastric cancer, initial staging is again by contrast-enhanced CT. If curative treatment is being considered, the use of endoscopic ultrasound (EUS) is helpful in determining the proximal and distal extent of the tumor, whereas 18F-FDG PET/CT has been shown to improve staging by detecting involved lymph nodes and metastatic disease, although it can be less accurate in mucinous and diffuse tumors [20].
In esophageal cancer, PET has the potential to change management in up to a third of patients [21,22], and is often incorporated into radiotherapy planning pathway [23,24]. The American College of Surgeons Oncology Group reported sensitivity and specificity of 18F-FDG-PET/CT scans to be 79% and 95%, respectively [18].
Magnetic resonance imaging (MRI) is currently not recommended for the routine imaging of esophageal or gastric cancer. However, with the recent advent of hybrid PET/MRI systems in clinical practice, there has been growing interest in MRI's ability as an assessment tool. MRI provides excellent soft-tissue contrast, and may demonstrate the esophageal wall layers and adjacent nodes. Physiological sequences (e.g., diffusion-weighted MRI) may also be included as part of the protocol. An initial staging 18F-FDG PET/MRI study with a pathology gold standard has been promising for nodal assessment with reported accuracy of 83% compared to 75% and 50% for EUS and CT, respectively [25].

A role for radiomics?
Radiomic approaches are showing promise for patient stratification. Radiomics exploit the data performed as part of the clinical management pathway. In terms of imaging, a number of parameters may be extracted and combined including standard descriptors (e.g., size, morphology, TNM (tumor, node, metastasis) stage); qualitative, semi-quantitative, or quantitative physiological parameters (e.g., contrast enhancement, diffusion characteristics, tracer uptake); and additional agnostic features which are otherwise 'invisible', with bioinformatic approaches. Of these, texture-based features have been investigated most commonly to date. Table 1 highlights some features that have been investigated in studies.
Radiomic signatures provide additional information predictive of underlying tumor biology and behavior. These signatures can be used alone or with other patient-related data (e.g., pathological data; genomic data) to improve tumor phenotyping, treatment response prediction and prognosis. Radiomic signatures may be obtained for all cross-sectional imaging modalities, including CT, PET, and MRI. figure 3 illustrates a typical radiomics pipeline. figure 4 demonstrates the process of tumor segmentation for a 18F-FDG PET image with a corresponding plot of standardized uptake value for the tumor. Initial studies in esophageal, esophago-gastric, and gastric cancer have shown promise for patient care.

18F-FDG PET radiomics
Nine 18F-FDG PET studies have been performed in esophageal and esophago-gastric cancer and are summarized in Table 2. As yet no studies have been performed for gastric cancer. Studies to date have focused on the prediction of response or prognosis in comparison to standard practice. Studies have found that various first, second and high-order features have been contributory to the assessment of response, differentiating between responders and non-responders (with greater heterogeneity in non-responders), as well as being predictive of complete response. Performance has been better than conventional parameters alone. Prognostication data remain conflicting.
In greater detail, five studies have investigated the prediction of response to therapy alone (n = 2); prognosis alone (n = 1) and the prediction of response to therapy and prognosis (n = 2) from pre-therapy imaging. One of the earliest studies by Tixier et al. showed in 41 patients that gray-level co-occurrence matrix (GLCM) homogeneity, GLCM entropy, gray-level sizezone matrix (GLSZM) size-zone variability and run length matrix (RLM) intensity variability differentiated non-responders, partial-responders, and complete-responders with sensitivities of 76%-92% [26]. Beukinga et al. showed in 97 patients that a clinical model including PET-derived gray-level run length (GLRL) long run low gray level emphasis and CT-derived run percentage had a higher area under the receiver operator curve (AUROC) compared to maximum standardized uptake value (SUV max ) in predicting therapy response [27].
In a study of 52 patients with squamous cancers, Nakajo et al. found that 18F-FDG PET/CT GLSZM intensity variability, and GLSZM size-zone variability, as well as standard volumetric parameters, such as metabolic tumor volume (MTV) and total lesion glycolysis (TLG), were predictors of tumor response but not of progression free or overall-survival

CT radiomics
Nine studies have investigated the ability of CT-derived heterogeneity parameters for classification, prediction of response and overall-survival in patients with esophageal or gastric cancer. Three studies have been performed for esophageal cancer in terms of prediction of response or prognosis (Table 3). These have found that greater heterogeneity is present in non-responders and those with poorer outcome.
The largest study in 49 patients found that histogram skewness, histogram kurtosis, GLSZM long-zone emphasis, and 2 Gabor transformed parameters MSA-54 and MSE-54, discriminated non-responders from responders using an artificial neural network-derived prediction model [35]. The two remaining studies have assessed prognostication. Ganeshan et al. found that lower histogram uniformity (with Gaussian filtration) from unenhanced CT images before start of treatment was an independent predictor for poorer overall-survival [36].Yip et al. analyzed contrast-enhanced images of 36 patients before and after treatment and found a significant decrease in histogram entropy and increase in uniformity (with Gaussian filtration) between the two time points. Higher post treatment entropy was associated with poorer overall-survival [37].
For gastric cancer (Table 4), three studies have assessed the potential of radiomic approaches for classification. Studies have found that first and second-order analysis in the contrastenhanced images may help in differentiation of lymphoma from gastrointestinal stromal cancer [38] or adenocarcinoma [39]. Another study in 107 patients found that arterial phase standard deviation and entropy were correlated with poorer differentiation [40]. For prognostication, Yoon et al. investigated 26 HER2 + gastric cancer patients before trastuzumab-treatment. In their analysis, they found GLCM contrast, variance, correlation and angular second moment (also known as energy or uniformity) were associated with a poorer survival [41]. In another study, Giganti et al. showed in 56 patients, that first-order energy, entropy, and skewness were significantly associated with a negative prognosis [42]. Giganti et al. also assessed pre chemotherapy texture features derived from the late arterial phase of 34 patients. They found entropy and compactness were higher and uniformity lower in responders [43]. No studies have assessed prognostication or response to therapy in gastric cancer.

MRI radiomics
To date there have been little data for MRI in this tumor group as MRI is not performed routinely in the clinical pathway. There have been some exploratory data of pre-therapeutic ADC-maps of gastric cancer (Table 5). Liu et al. found that first-order statistics skewness may differ from node positive to node negative patients, and are associated with pathological characteristics including perineural and vascular invasion [44][45][46][47]. However, no studies so far have investigated prognostication or response assessment.

Discussion
To date 22 imaging studies have been published investigating radiomic approaches in esophageal, esophagogastric, and gastric cancer, predominantly focused on texture analysis. Preliminary data for esophageal and esophago-gastric cancer suggest that there is potential for radiomic approaches in improving patient stratification for therapy. Eight 18F-FDG PET studies investigated the feasibility of heterogeneity analysis for response prediction (four studies with pre-therapy imaging only). Among the most often reported significant feature was GLCM entropy, reflecting the local randomness (irregularity) within the image, and where low GLCM entropy represents a more homogeneous texture. The reported accuracy for successful classification of therapy response ranged from 0.7 to 1.0 (AUC).
Nearly all published studies incorporated "classical" PET parameters e.g., SUV max , total lesion glycolysis and metabolic tumor volume into predictive models. In general radiomic parameters contributed to predictive models and provided additional information to standard parameters. Three CT studies of esophageal cancer have also suggested that greater tumoral heterogeneity is associated with poor response and outcome.
PET-studies investigating texture features as a prognosticator were more mixed. Only two studies found associations with overall-survival. The CT and MRI data for gastric cancer were also varied. Two studies found several features to be associated with survival time, however, for some parameters, e.g., histogram entropy and energy, it was surprising to find both parameters to be associated in the same direction given what they represent mathematically.
A challenge for interpretation of studies to date is the use of retrospective datasets with different imaging techniques across different scanners and/or institutions; different methodologies for feature selection; the focus on different feature sets; the lack of transparency in methodology with the use of different in-house software; as well as varying statistical and bioinformatics approaches for data analysis and interpretation. This has been highlighted by researchers in the field [48].
Moving forward in the context of esophageal and esophago-gastric cancer, it is important to improve our data quality. Planned prospective studies incorporating quality control is a step in the right direction to improving data curation and ensuring prediction models are fit for purpose and fulfill the promise of radiomics for improving patient stratification.  Example of tumor segmentation for extraction of radiomic features from an axial PET image. In the right image the corresponding standardized uptake values for the region-ofinterest is displayed.