Imaging biomarkers in upper gastrointestinal cancers.

In parallel with the increasingly widespread availability of high performance imaging platforms and recent progresses in pathobiological characterisation and treatment of gastrointestinal malignancies, imaging biomarkers have become a major research topic due to their potential to provide additional quantitative information to conventional imaging modalities that can improve accuracy at staging and follow-up, predict outcome, and guide treatment planning in an individualised manner. The aim of this review is to briefly examine the status of current knowledge about imaging biomarkers in the field of upper gastrointestinal cancers, highlighting their potential applications and future perspectives in patient management from diagnosis onwards.

introduction

The term “biomarker” refers to a characteristic that is objectively measured and evaluated as an indicator of normal biologic processes, pathogenic processes, or biological responses to therapeutic interventions.[1-3] Biomarkers can be derived directly from human biological samples (including blood or other biological fluids, tissues or cells) or biosignals detected by diagnostic imaging techniques (such as multidetector CT, MRI, ultrasound or PET), providing anatomical and/or functional parameters that correlate with histopathological or molecular data.[4-7] Imaging biomarkers can be either quantitative (e.g. lesion diameter or volume, CT density, MR signal intensity, radiomics features, and any other biomarker whose magnitude can be expressed as a quantity value) or qualitative (e.g. pathological grading systems that can be expressed as ordinal rather than continuous quantitative data, such as clinical TNM staging, BI-RADS, LI-RADS, PI-RADS, C-RADS, and so on).[8,9] Compared to “real” biological samples, imaging biomarkers have the advantage of being noninvasive (or minimally invasive) and easily obtainable, making them attractive for repeated testing and follow-up. In addition, while conventional biopsy captures only a fraction of tumour cells that may not be representative of the entire tumour, imaging biomarkers can capture the whole tumour heterogeneity, thus playing a potentially critical role into guiding biopsy or whenever biopsy is unobtainable.[10-12]

Due to the incremental diagnostic and prognostic value that imaging biomarkers can provide over conventional modalities (which, however, remain indispensable for the routine diagnostic workup of cancer patients), a great interest has arisen in the scientific and medical community in searching for biomarkers and building biobanks that may lead to more effective, patient-specific treatment. In this perspective, extensive research has recently been made in the field of radiomics, a discipline dealing with high-throughput extraction and analysis of quantitative features from diagnostic images that cannot usually be derived from visual assessment (including shape/size-, histogram- or filter-based texture analysis) and form the quantitative imaging expression of pathological processes that are also expressed by other “omics”, such as genomics, transcriptomics, metabolomics and proteomics.[5,8]

In this paper we will give a brief, yet hopefully explanatory overview of the current potential role of imaging biomarkers in the management of upper gastrointestinal cancers, with oropharyngeal, oesophageal, and gastric cancers as paradigm diseases.

oropharyngeal cancer

Head and neck cancer is the seventh most frequent cancer and the ninth most frequent cause of cancer-related death, with a worldwide incidence of 550,000 cases and 380,000 deaths annually, of which 90% are squamous cell carcinomas (SCC) in the adult population.[13,14] Most oropharyngeal cancers are diagnosed at a locally advanced stage, usually requiring multimodality treatment (including surgery, chemotherapy and/or radiation therapy), and the rate of locoregional recurrence is as high as 25–50% depending on the lesion site, making recurrence risk the greatest hurdle in improving survival rate. Moreover, tumour biology (with particular reference to human papillomavirus [HPV] positivity) plays a pivotal role in defining the prognosis and optimal therapeutic strategy for patients with head and neck SCC, with HPV+oropharyngeal SCC being associated with a better overall survival. Therefore, a comprehensive assessment of disease extension (including primary tumour site and locoregional lymph nodes) at baseline and after treatment is mandatory, also in consideration that oropharyngeal and other head and neck SCC recurrences tend to occur at a locoregional level within the same volume of treatment.[15-17]

Although conventional multidetector CT and MR imaging performed with state-of-the-art equipment is the mainstay for diagnosis, staging and follow-up of patients with oropharyngeal SCC, morphologic imaging can be limited in detecting subtle disease or differentiating disease recurrence from benign post-treatment changes (e.g. after surgery and/or radiation therapy) that may obscure viable tumour tissue or mimic disease, and usually cannot provide per se any insight about prognosis, likelihood of early and/or sustained treatment response, or eligibility to patient-tailored target therapies.[14,18-20] On the other hand, both standard CT and MRI data and those derived from more complex implementations of cross-sectional imaging can be joined to build specific radiomic signatures that can lead to improved patient stratification and potentially more personalised treatment[8,21] (Table 1). In this latter setting, a recent large multicentre study by Leijenaar et al combining 902 radiomic features from 778 patients with oropharyngeal SCC showed that CT-based radiomics can predict HPV+ (p16) status from multivariable models obtained from standard pre-treatment CT images. More specifically, a model was developed on all training data (Mall) and on an artefact-free subset of training data (Mno art) to assess the impact of CT artefacts in predicting HPV (p16), and a consistent and significant split was observed between survival curves with HPV status determined by p16 [p = 0.007, hazard ratio (HR) 0.46], Mall (p = 0.036, HR 0.55) and Mno art (p = 0.027, HR 0.49), along with an area under the receiver operator curve ranging between 0.70 and 0.80 for both Mall and Mno art.[22] In addition, evidence is growing that a radiomic approach can provide useful prognostic quantitative information in patients undergoing chemo-radiation therapy (CRT), such as primary tumour volume (which correlates with treatment outcome better than T or N stages),[23] CT texture analysis (which can predict local failure after CRT),[24] and 18F-FDG PET texture as a predictor of post-CRT patient survival in addition to clinical variables.[25]

Table 1. 

Summary of the features and main findings of the studies on imaging biomarkers in head and neck cancer (with particular reference to oropharyngeal cancer) cited in the text. Case reports and general reviews articles are not included. References are sorted in the order in which they appear in the text

Reference	Tumour type and location	Imaging features	Endpoints	Main findings
van der Hoorn et al.[20]	Head and neck SCC (oral cavity, pharynx, larynx)a	Anatomical MRI, ADC	Diagnostic accuracy for treatment response evaluation	Pooled analysis of anatomical MRI of the primary site showed a sensitivity of 84% (95% confidence interval 72÷92%) and specificity of 82% (71÷89%) ADC of the primary site showed a pooled sensitivity of 89% (74÷96%) and specificity of 86% (69÷94%)
Leijenaar et al.[22]	HPV+ oropharyngeal SCC	Radiomic signatures (Mall, Mno art) for HPV (p16) prediction from standard pre-RT/CRT CT imaging	Diagnostic accuracy, OS	AUC for Mall and Mno art as high as 0.70÷0.80 Consistent and significant split between survival curves with HPV status determined by p16 (p = 0.007, HR 0.46), Mall (p = 0.036, HR 0.55) and Mno art (p = 0.027, HR 0.49)
Strongin et al.[23]	Locally advanced SCC (oropharynx, hypopharynx and larynx)	Primary tumour volume (TV) before CRT	PFS, OS	Better 5-year prognosis with primary TV <35 cm3 (71% vs 43%, p = 0.01) Longer PFS and OS with primary TV <35 cm3 (61% vs 33%, p = 0.022 and 84 vs 41%, p < 0.001, respectively) Primary TV was best predictor of recurrence (HR 4.7, p = 0.001) and survival (HR 10.0, p < 0.001) on multivariate analysis TV of SCC with locoregional failure was larger compared with those without locoregional failure and those recurring as distant metastases (p < 0.03)
Kuno et al.[24]	Head and neck SCC (oropharynx, larynx, hypopharynx and oral cavity)	CT texture parameters	Local recurrence after CRT	3 histogram features (geometric mean, HR = 4.68, p = 0.026; harmonic mean, HR = 8.61, p = 0.004; fourth moment, HR = 4.56, p = 0.048) and four gray-level run-length features (short-run emphasis, HR = 3.75, p = 0.044; gray-level nonuniformity, HR = 5.72, p = 0.004; run-length nonuniformity, HR 4.15, p = 0.043; short-run low gray-level emphasis, HR = 5.94, p = 0.035) were significant predictors of outcome after adjusting for clinical variables
Feliciani et al.[25]	Head and neck SCC(oral cavity, oropharynx, hypopharynx, nasopharynx and larynx)	Pre-CRT 18F-FDG PET texture features for prediction of treatment failure	Local control rate, PFS, OS	Low-intensity long-run emphasis (LILRE) was a significant predictor of outcome regardless of clinical variables (HR <0.001, p = 0.001) Better local failure prediction using multivariate model based on imaging biomarkers than that based on clinical variables alone (c-index 0.76 vs 0.65)
Holzapfel et al.[26]	SCC cervical nodal metastases	ADC	Diagnostic accuracy for differentiating between benign and metastatic lymph nodes	ADC values of malignant lymph nodes were significantly lower than ADC values of benign lymph nodes (0.78 ± 0.09 vs 1.24 ± 0.16×10−3 mm2/s, p < 0.05) 94.3% of lesions were correctly classified as benign or malignant using a threshold ADC value of 1.02 × 10−3 mm2/s (AUC 0.975)
Jin et al.[27]	Nasopharyngeal SCC cervical nodal metastases*	ADC	Diagnostic accuracy for differentiating between benign and metastatic lymph nodes	Benign lymph nodes had higher ADC values than metastatic nodes (1.110 ± 0.202 vs 0.878 ± 0.159×10−3 mm2/s, p < 0.05) An ADC cut-off value of 0.924 × 10−3 mm2/s yielded a sensitivity, specificity, PPV, NPV and accuracy of 83.56%, 82.74%, 80.79%, 85.28%, and 82.80%, respectively for differentiating metastatic from benign lymph nodes (AUC 0.851)
Hwang et al.[28]	Head and neck SCC (oral cavity, oropharynx, sinonasal cavity, nasopharynx, hypopharynx, external auditory canal)	ADC	Diagnostic accuracy for differentiating tumour recurrence from post-treatment (surgery and/or chemo/radiotherapy) changes	Significantly lower ADC1000 in recurrent tumour than in post-treatment changes (p < 0.001) Significantly higher ADCratio in recurrent tumour than that in post-treatment changes (73.5 ±7.2% vs 56.9 ±8.8%, p < 0.001) ADCratio was the only independently differentiating variable (p = 0.024) An ADCratio cut-off value of 62.6% yielded a sensitivity, specificity and accuracy of 95.0%, 69.2% and 84.8%, respectively
Ryoo et al.[29]	Head and neck SCC (oral cavity, oropharynx, nasopharynx, supraglottic larynx, maxillary sinus)	ADC	Diagnostic accuracy for predicting response to induction chemotherapy	Good responders had significantly lower ADC2000 values than poor responders (0.62 ± 0.14 vs 0.76 ± 0.15×10⁻³mm²/s, p = 0.02) Mean ADC2000 was a significant predictor of response to induction chemotherapy at multiple logistic regression analysis (p = 0.04)
Preda et al.[30]	Head and neck SCC	SUV, ADC	DFS	Patients with SUVmaxT/B ≥5.75 had an overall worse prognosis (p = 0.003) Lymph node status- and diameter-adjusted SUVmaxT/B and ADCmin were significant predictors of DFS (HR = 10.37 and 3.26 for SUVmaxT/B ≥5.75 and ADCmin ≥0.58 ×10⁻³ mm²/s, respectively) In patients with SUVmaxT/B ≥5.75, high ADCmin was a significant predictor of worse prognosis (adjusted HR = 3.11)
Surov et al.[31]	Head and neck SCC	SUV, ADC	Tumour proliferation indexes (Ki67, cell count, total nucleic area, average nucleic area)	ADCmean correlated with Ki67 level (r = −0.728, p = 0.011) and total nucleic area (r = −0.691, p = 0.019) ADCmax correlated with Ki67 level (r = −0.633, p = 0.036) Combined parameter SUVmax/ADCmin correlated with average nucleic area (r = 0.627, p = 0.039)
Dong et al.[34]	Head and neck SCC	DCE-MRI (Ktrans, Ve, iAUC)	Tumour grading	Higher Ktrans, Ve and iAUC in poorly differentiated than in well differentiated SCC (p < 0.001, 0.013 and <0.001, respectively) Ktrans had the greatest diagnostic significance, with a cut-off value of 0.270 min−1 yielding a Youden index, sensitivity and specificity of 0.859, 95.0% and 90.9%, respectively
Chen et al.[35]	Head and neck cancer (nasopharynx, oropharynx, tongue base, larynx)	DCE-MRI (Ktrans, Ve, kep)	Differentiation among tumours, metastatic nodes and normal tissue	Ktrans and Ve of normal tissue (0.159 ± 0.087 min−1 and 0.229 ± 0.146) were lower than those of nodes (0.332 ± 0.149 min−1 and 0.408 ± 0.124, p < 0.001) and primary tumours (0.251 ± 0.066 min−1 and 0.344 ± 0.081, p < 0.001), respectively kep values of primary tumours (0.621 ± 0.195 min−1) were significantly higher than those of nodes (0.429 ± 0.206 min−1, p = 0.001) and normal tissue (0.420 ± 0.170 min−1, p = 0.002)
Baker et al.[36]	Head and neck SCC murine xenografts	R1 and R2* relaxation, functional MRI, relative 18F-FDG uptake	Association with acquired resistance to targeted EGFR therapy within size-matched EGFR TKI-resistant CAL 27 (CALR) and sensitive (CALS) tumour xenografts in vivo	Lower baseline R2* (58 ± 2 vs 70 ± 2 s−1), hyperoxia-induced ΔR2* (−4 ± 1 vs −11 ± 2 s−1) and volume transfer constant Ktrans values in CALR than in CALS tumours (p < 0.01), associated with significantly lower Hoechst 33342 uptake and greater pimonidazole-adduct formation Higher relative 18F-FDG uptake in the CALR cohort (48%, p < 0.05), associated with significantly greater Glut-1 expression
Zhong et al.[37]	Lymph node metastases from head and neck SCC (larynx, tongue, nasopharynx, floor of mouth, nasal cavity, oropharynx, gingiva)	Preoperative DWI and perfusion CT	Diagnostic accuracy for differentiation between metastatic and benign lymph nodes	Lower mean ADC in metastatic than in benign nodes (0.849 ± 0.111 vs 1.443 ± 0.406×10⁻³ mm²/s, p = 0.011) Higher BF (114.62 ± 14.26 vs 67.82 ± 13.84 ml/100 g/min, p = 0.002) and lower MTT (5.56 ± 0.39 vs 9.46 ± 3.23 s, p = 0.002) in metastatic than in benign nodes An ADC cut-off value of 0.960 × 10⁻³ mm²/s yielded a sensitivity, specificity, PPV, NPV and diagnostic accuracy of 89.58%, 76.47%, 91.48%, 72.22% and 86.15%, respectively ADC had better diagnostic accuracy than perfusion CT in differentiating metastatic from benign lymph nodes (AUC 0.320 vs 0.605, p < 0.001)
Lam et al.[40]	Head and neck SCC (primary and lymph node metastases)	DECT image quality	SNR, attenuation difference between tumour and muscles, lesion CNR	65keV monochromatic images yielded optimal signal-to-noise ratio for all tissues (p < 0.0001) 40keV monochromatic images yielded best tumour attenuation (p < 0.0001), attenuation difference between tumour and muscles (p = 0.03), and lesion CNR (p < 0.0001)
Tawfik et al.[43]	Metastatic SCC cervical lymph nodes	DECT-derived iodine content and iodine overlay	Differentiation among normal, inflammatory and metastatic SCC lymph nodes	Iodine content was significantly lower for metastatic lymph nodes (2.34 ± 0.45 mg ml−1) than for normal (2.86 ± 0.37 mg ml−1) and inflammatory (3.53 ± 0.56 mg ml−1) lymph nodes (p < 0.0001) An iodine content cut-off <2.85 mg ml−1 to diagnose nodal metastases had 85% sensitivity and 87.5% specificity (AUC 0.923) Iodine overlay was significantly lower for metastatic lymph nodes (47 ± 11.6 HU) than normal (57.4 ± 8.2 HU) and inflammatory nodes (69.3 ± 11.5 HU) (p < 0.0001) An iodine overlay cut-off <57.5 HU to diagnose nodal metastases had 90% sensitivity and 78% specificity (AUC 0.896)
Foust et al. 2018[44]	Oropharyngeal SCC with nodal metastases	DECT-derived iodine content and iodine spectral attenuation slope	Diagnostic accuracy for differentiation between metastatic and nonmetastatic lymph nodes	Iodine content was significantly lower in metastatic than in nonmetastatic nodes (0.96 ± 0.28 vs 1.65 ± 0.38 mg ml−1, p = 0.002) Iodine spectral attenuation slope was significantly lower in metastatic than in nonmetastatic nodes (1.33 ± 0.49 vs 1.91 ± 0.64 mg ml−1, p = 0.015) A nodal iodine threshold ≤1.3 mg ml−1 showed a sensitivity of 84.6% and a specificity of 75.0% (AUC 0.839, p < 0.0001) A nodal spectral attenuation slope threshold ≤1.95 showed a sensitivity of 92.3% and specificity of 50.0% (AUC 0.68, p = 0.049)
Rizzo et al.[45]	Metastatic lymph nodes from various primary cancers (lung, gynaecological)*	DECT-derived material decomposition maps and intranodal attenuation (HU) distribution	Differentiation between metastatic and nonmetastatic lymph nodes based on spectral structure	Iodine-fat decomposition values were lower in metastatic than in nonmetastatic lymph nodes (3.50 ± 1.02 vs 4.36 ± 1.06 mg ml−1, p < 0.001) Iodine-water decomposition values were lower in metastatic than in nonmetastatic lymph nodes (22.4 ± 9.0 vs 31.6 ± 9.8×100 µg ml−1, p < 0.001) HU distribution showed a significant gradient from centre to periphery within non-metastatic lymph nodes (up to 20–30% from the centre at 40keV, p < 0.035) Metastatic lymph nodes showed no significant HU gradient

ADC, apparent diffusion coefficient; AUC, area under the ROC curve; CNR, contrast-to-noise ratio; DFS, disease free survival; HR, hazard ratio; NPV, negative predictive value; OS, overall survival; PFS, progression free survival; PPV, positive predictive value; SNR, signal-to-noise ratio; SUV, standardised uptake value.

Meta-analysis article. (*) References 27 and 45 were included due to the general value of their respective findings, although not specific to oropharyngeal cancer.

Among MRI techniques, diffusion-weighted imaging (DWI) is a powerful quantitative biomarker that can identify foci of residual disease as areas of restricted diffusivity of water molecules, due to cancer cells being densely packed with a relatively high nuclear-cytoplasmic ratio compared with non-neoplastic tissues. This allows overcoming potential confounding factors such as diffuse oedema and contrast enhancement as a result of inflammation and/or post-radiation sequelae, or faint or no enhancement in poorly vascularised and/or highly necrotic tumours, resulting in increased sensitivity in detection of primary oropharyngeal SCC.[14,20] In a recent large meta-analysis, DWI-based apparent diffusion coefficient (ADC) quantitative assessment outperformed anatomical MRI for detection of primary tumour site with a pooled sensitivity and specificity of 89% and 86% vs 84% and 82%, respectively.[20] Lower ADC values have also been found in malignant than in benign lymph nodes sized down to 5–10 mm, and Holzapfel et al were able to correctly classify cervical lymph nodes as benign or malignant using an ADC threshold of 1.02 × 10−3 mm2/s in 94.3% of cases.[14,26,27] Furthermore, Hwang et al assessed the performance of standard (b = 1000 s/mm2) and high b-value (2000 s/mm2) DWI in differentiating between recurrent tumour and post-treatment changes, and found that the ratio between ADC2000 and ADC1000 was significantly higher in the former than in the latter (73.5% vs 56.9%, p < 0.001), with a cut-off value of 62.6% leading to a sensitivity, specificity, and diagnostic accuracy of 95.0%, 69.2%, and 84.8%, respectively.[28] Of interest, patients with a good response to induction neoadjuvant chemotherapy were found to have lower ADC2000 values along with smaller tumour volume than poor responders, suggesting a potential role of high ADC values in predicting tumour response.[29] In terms of risk stratification, it has been reported that the combination of poor prognostic factors such as high 8F-FDG PET-CT-derived SUVmaxT/B and high ADCmin values improves the prognostic role of the two separate parameters (with high ADCmin identifying SCC patients with worse prognosis among those with high SUVmaxT/B).[30] Moreover, a significant association was observed between histopathological markers of tumour proliferation and ADC and SUV values calculated at simultaneous 18F-FDG-PET/MRI, with mean ADC being negatively correlated with Ki67 level (r = −0.728, p = 0.011) and total nucleic area (r = −0.691, p = 0.019), and combined SUVmax/ADCmin positively correlated with average nucleic area (r = 0.627, p = 0.039).[31]

Other functional imaging modalities such as dynamic contrast enhancement (DCE)-MRI, CT perfusion imaging and dual energy CT (DECT) can enable a quantitative evaluation of tumour vascularity that can be useful in the pre- and post-treatment setting.[14,32] In general, highly vascular SCC tend to have a better treatment response compared to less vascular ones due to their greater sensitivity to radiation and better delivery of chemotherapeutic drugs, yet they may have poorer outcome due to a higher metastatic potential.[14,33] In terms of grading, quantitative DCE-MRI parameters such as the blood volume transfer constant (Ktrans), the extracellular volume fraction (Ve), and the initial area under the curve (iAUC) were shown to be higher in poorly than in well-differentiated SCC, with Ktrans having the greatest diagnostic significance at a cut-off value of 0.270 min−1 and an overall diagnostic sensitivity and specificity of 95.0 and 90.9%.[34] Chen et al found that DCE-MRI parameters [Ktrans, Ve and the rate constant of contrast transfer (Kep)] allowed to discriminate among tumour, metastatic lymph nodes and normal tissue, with Ktrans and Ve values of normal tissue being significantly lower from those of nodes (p < 0.001) and primary tumours (p < 0.001), whereas Kep values of primary tumours were significantly higher from those of nodes (p = 0.001) and normal tissue (p = 0.002), respectively.[35] Moreover, Baker et al reported that DCE-MRI can predict tumour response to target therapies with EGFR-targeted tyrosine kinase inhibitors, in conjunction with 18F-FDG PET.[36]

Similarly, perfusion CT can quantify the vascularity of primary head-and-neck SCC by applying specific bio-mathematical models of tumour neoangiogenesis (Figure 1). Evidence exists that a) better treatment response can be usually be predicted at perfusion CT before the beginning of therapy in more perfused tumours, and b) early perfusion response to CRT can be assessed before the onset of morphological changes at conventional morphological imaging.[37,38] On the other hand, DECT is emerging as an increasingly widespread imaging technique that can join the fast imaging time, high spatial resolution, and excellent morphological detail of multidetector CT with the possibility to collect functional data via material decomposition, allowing to calculate iodine content as a biomarker of tumour vascularity with a potentially lower overall radiation and contrast dose compared with standard single-energy CT imaging[39] (Figure 2). DECT can be effective in improving the detection and local staging of primary tumours, with particular reference to N-staging, which is often limited due to small and/or morphologically indeterminate lymph nodes leading to inconclusive findings at conventional anatomical imaging.[38,40-42] To this regard, Tawfik et al reported significant differences in DECT-derived iodine content and overlay among normal, inflammatory and metastatic SCC lymph nodes.[43] Moreover, DECT showed lower iodine content and spectral attenuation slope in metastatic oropharyngeal SCC than in non-neoplastic cervical lymph nodes, in line with the finding of a different intranodal iodine distribution that reflects pathological structure.[44,45]

Figure 1.

An 83-year-old male with locally advanced, partially necrotic gingival SCC as displayed on standard morphological CT image (a). Quantitative CT perfusion analysis revealed increased blood flow (BF) (b), blood volume (BV) (c), permeability surface (PS) product (d), and reduced mean transit time (MTT) (e) values in the viable portion of the tumour compared with contralateral normal tissue (Ref) taken as reference. Note lack of perfusion inside the necrotic portion of the tumour as visually depicted on colour-coded perfusion maps.

Figure 2.

A 51-year-old female with small keratinising SCC of the left retromolar trigone (arrow on DECT monochromatic 55keV image, (a). Material decomposition iodine/water image (b) improves lesion conspicuity by boosting iodine signal and reveals high lesion vascularisation (measured mean iodine concentration of 32 × 100 µg ml−1), whereas virtual precontrast image (c) shows no abnormal density or beam hardening artefacts at the lesion site. Avid lesion enhancement was confirmed by spectral curve analysis (d). All images had been obtained at the same anatomic level from a single contrast-enhanced DECT acquisition.

Overall, the integration of quantitative image biomarkers in the diagnostic workup of patients with oropharyngeal cancer and other tumours of the uppermost portion of the digestive tract looks promising in an effort to optimise lesion detection and staging, either at primary sites (especially in case of suspected disease recurrence, when diagnosis by conventional imaging is often hampered by gross post-treatment sequelae) and at lymph node levels (possibly overcoming the poor performance of standard morphological imaging in differentiating smaller benign from metastatic lymph nodes and guiding the choice towards more or less aggressive lymph node resection). Moreover, the correlation between imaging biomarkers and the biological features of tumours (related e.g. to lesion aggressiveness) may be of great value to improve patient selection for the best possible treatment in an individualised fashion.

oesophageal and gastric cancer

Oesophageal cancer is the eighth most common cancer and the sixth most common cause of cancer-related deaths worldwide, accounting for more than 509,000 deaths in 2018.[46] While initial staging of oesophago-gastric cancer is usually performed with contrast-enhanced CT, multimodality imaging with endoscopic ultrasound and 18F-FDG PET is useful to define local disease extent and rule out metastatic spread in patients being considered for curative treatment, in an effort to reduce futile surgery.[47] MRI is a less routinely used, yet valuable tool that can reveal the detailed anatomic structure of the oesophagus, differentiating among the various wall layers for proper local staging and providing an accurate assessment of lesion length before and after radiation therapy.[48,49]

Several imaging biomarkers have been identified that can help predict response of oesophageal cancer to CRT (Table 2). In this latter setting, Aoyagi et al reported that ADC values were significantly higher in patients with better survival rates,[50] whereas in a recent meta-analysis the variation of ADC before and after treatment and post-treatment ADC were found to have good performance for evaluating CRT response.[51] Furthermore, van Rossum et al showed that a treatment-induced relative increase in ADC during the first 2–3 weeks of neoadjuvant CRT for oesophageal cancer was highly predictive of histopathologic response,[52] and also AUC changes at DCE-MRI were predictive of complete pathological response to neoadjuvant CRT at a threshold of −24.6% with sensitivity, specificity, positive and negative predictive values of 83%, 88%, 71% and 93%, respectively.[53]

Table 2.

Summary of the features and main findings of the studies on imaging biomarkers in oesophageal cancer cited in the text. Case reports and general reviews articles are not included. References are sorted in the order in which they appear in the text

Reference	Tumour type and location	Imaging features	Endpoints	Main findings
Ge et al.[46]	Oesophageal cancer	Pre- and post-CRT DECT-derived normalised iodine concentration (NIC) and normalised CT (NCT) values	Comparison of pre- and post-CRT NIC and NCT values obtained in the arterial (NIC-A, NCT-A) and portal venous phases (NIC-V, NCT-V) in good and poor responders	Post-CRT NIC-A, NIC-V, NCT-A and NCT-V values were significantly lower than before CRT in good responders (0.23 ± 0.05 vs 0.28 ± 0.05, 0.49 ± 0.06 vs 0.54 ± 0.06, 0.27 ± 0.04 vs 0.34 ± 0.06, 0.50 ± 0.04 vs 0.57 ± 0.80, respectively; p < 0.05) Post-CRT NIC-A, NIC-V, NCT-A and NCT-V values were significantly lower in good than in poor responders (0.23 ± 0.05 vs 0.29 ± 0.06, 0.49 ± 0.06 vs 0.51 ± 0.06, 0.27 ± 0.04 vs 0.31 ± 0.04, 0.50 ± 0.04 vs 0.54 ± 0.05, respectively; p < 0.05) Post-CRT NIC-V and NCT-V values in poor responders were significantly lower than before CRT (0.51 ± 0.06 vs 0.56 ± 0.06 and 0.54 ± 0.05 vs 0.59 ± 0.04, respectively; p < 0.05)
Guo et al.[49]	Oesophageal cancer (SCC, adenocarcinoma, small cell carcinoma)	CT, ADC	Diagnostic accuracy before treatment, outcome prediction during and after radiotherapy in CR and PR patients	Significantly greater difference in length of oesophageal lesions measured at CT (1.15 ± 0.59 cm) than at DWI (from 0.19 ± 0.36 cm at b = 600 mm2/s to 0.84 ± 0.53 cm at b = 1000 mm2/s) compared with pathological specimens (p < 0.05) Higher diagnostic rate at DWI than at CT (98.72% vs 88.46%, p = 0.022) After radiotherapy, the clinical control rate and 3-year survival rate in the high ADC value group were higher than in the low ADC value group (90.24% vs 67.57%, p = 0.028 and 58.54% vs 32.43%, p = 0.037, respectively) In the second week during radiotherapy and at the end of radiotherapy, the ADC values in the CR group were significantly higher than in the PR group (p < 0.05) ADC values measured in the second week during radiotherapy and at the end of radiotherapy had an AUC of 0.776 and 0.935, respectively for predicting the CR rate of radiotherapy
Aoyagi et al.[50]	Oesophageal SCC	Pre-CRT ADC	Prediction of CRT response or prognosis	Higher ADC values were associated with CRT response (1.27 ± 0.17 vs 0.92 ± 0.28×10−3 mm2/s, p < 0.01) ADC values higher than the average ADC of oesophageal cancer tissue (1.10 × 10−3 mm2/s) were associated with a higher survival rate (median OS according to ADC of 309 vs 197 days, p = 0.02)
Cheng et al.[51]	Oesophageal cancer	ADC [ADC variation before and after CRT (∆ADC), post-ADC]	Prediction of early response to CRT	∆ADC showed a pooled sensitivity, specificity, diagnostic odds ratio, and AUC of 93%, 85%, 78 and 0.91 Post-ADC showed a pooled sensitivity, specificity, diagnostic odds ratio, and AUC of 75%, 90%, 26 and 0.85
van Rossum et al.[52]	Oesophageal cancer (adenocarcinoma, SCC)	ADC [ADC variation during treatment (∆ADCduring)]	Prediction of pathologic response to neoadjuvant CRT	ΔADCduring was significantly higher in patients with vs without pathologic CR (34.6±10.7% vs 14.0±13.1%, p = 0.016) and in good vs poor responders (30.5±8.3% vs 9.5±12.5%, p = 0.002) ΔADCduring was predictive of residual cancer at a threshold of 29% (sensitivity, specificity, PPV and NPV of 100%, 75%, 94% and 100%, respectively) ΔADCduring was predictive of poor pathologic response at a threshold of 21% (sensitivity, specificity, PPV and NPV of 82%, 100%, 100% and 80%, respectively)
Heethuis et al.[53]	Oesophageal cancer (SCC, adenocarcinoma, adenosquamous carcinoma)	DCE-MRI [change of AUC before neoadjuvant CRT (AUCpre), at 2–3 weeks during neoadjuvant CRT (AUCper), and after neoadjuvant CRT before surgery (AUCpost)]	Prediction of pathologic response to neoadjuvant CRT	The difference between AUCper and AUCpre was most predictive for good response at a threshold of 22.7% (sensitivity, specificity, PPV and NPV of 92%, 77%, 79% and 91%, respectively) The difference between AUCpost and AUCpre was most predictive for pathologic complete response at a threshold of −24.6% (sensitivity, specificity, PPV and NPV of 83%, 88%, 71% and 93%, respectively)
Shen et al.[54]	Lymph node metastases from oesophageal cancer	CT radiomics features	Prediction of lymph node metastasis status in the preoperative setting	Significant association between radiomics signature and lymph node metastasis (p < 0.001), with an AUC of 0.806 and 0.771 in the training and validation cohort, respectively Good discrimination of the predictive nomogram model (Harrell's Concordance Index of 0.768 and 0.754 in the training and validation cohort, respectively)
Hou et al.[55]	Oesophageal carcinoma	CT radiomics features	Prediction of CRT response	Five radiomics features (Histogram2D_skewness, Histogram2D_kurtosis, GLSZM2D_LZE, Gabor2D_MSA-54, Gabor2D_MSE-54) discriminated nonresponders from responders (AUCs 0.686÷0.727) Two features (Histogram2D_skewness and Histogram2D_kurtosis) allowed differentiating SDs from PRs (p = 0.015 and p = 0.039, respectively) One feature (Histogram2D_skewness) allowed differentiating SDs from CRs (p = 0.027) Both ANN and SVM models had high accuracy for potentially predicting treatment response (ANN 0.972, SVM 0.891)
Hou et al.[56]	Oesophageal SCC	MRI radiomics features (T2w, SPAIR T2w)	Prediction of CRT response	CRs vs SDs, PRs vs SDs, and responders (CRs and PRs) vs nonresponders (SDs) could be differentiated by 26, 17, and 33 features (T2w: 11/11/15, SPAIR T2w: 15/6/18), respectively Prediction models (ANN and SVM) based on features extracted from SPAIR T2w sequences showed higher accuracy than those derived from T2w sequences (SVM 0.929 vs 0.893 and ANN 0.883 vs 0.861, respectively)
Tan et al.[57]	Lymph node metastases from resectable oesophageal SCC	CT radiomics features	Diagnostic accuracy	The radiomics signature including five features was significantly associated with lymph node metastasis The radiomics nomogram [incorporating the signature and CT-reported lymph node status (i.e. size criteria)] distinguished lymph node metastasis with an AUC of 0.758 and 0.773 in the training and test set, respectively Discrimination of the radiomics nomogram exceeded that of size criteria alone in both the training (p < 0.001) and test sets (p = 0.005) Integrated discrimination improvement (IDI) and categorical net reclassification improvement (NRI) showed significant improvement in prognostic value when the radiomics signature was added to size criteria in the test set (IDI 17.3%, p < 0.001; categorical NRI 52.3%, p < 0.001)

CR, complete response; PR, partial response; SD, stable disease.

In parallel, a comparative study of DECT and standardised iodine concentrations in patients with oesophageal cancer before and after CRT showed a significant reduction of tumour normalised iodine concentration on iodine maps in responders, allowing to functionally evaluate treatment response and prognosis in terms of reduced tumour iodine uptake.[46] Radiomics approaches have also recently been proposed for preoperative prediction of lymph node metastasis[54] and tumour response to CRT in patients with oesophageal cancer,[55,56] and of note, it has emerged that radiomics features can outperform size criteria in discriminating lymph node metastasis in resectable oesophageal SCC.[57]

Gastric cancer is the fourth most common malignancy and the second most common cause of cancer-related death worldwide (738,000 deaths annually), whose prognosis and survival rate are poor in advanced stage disease and for which only surgical resection of the primary tumour can be curative.[58,59] In light of this, quantitative imaging represents a promising area for gastric cancer research and management, and in particular, ADC and contrast-enhanced CT-based texture features are noninvasive, quantitative imaging biomarkers that hold promise in the evaluation of tumour aggressiveness, treatment response and prognosis (Table 3).[47,60,61] CT texture analysis can predict histopathological characteristics of gastric cancers (such as differentiation degree, Lauren classification and vascular invasion status),[62] the expression level of immunohistochemical biomarkers (including E-cadherin, Ki67, and VEGFR2),[63] and even occult peritoneal carcinomatosis in patients with advanced disease[64] (Figure 3). Radiomic feature extraction is another area of active research that has yielded promising results, especially in terms of prognostic assessment.[47,65-67]

Table 3. 

Summary of the features and main findings of the studies on imaging biomarkers in gastric cancer cited in the text. Case reports and general reviews articles are not included. References are sorted in the order in which they appear in the text

Reference	Tumour type and location	Imaging features	Endpoints	Main findings
Giganti et al.[58]	Resectable gastric cancer (adenocarcinoma, signet-ring cell carcinoma)	ADC	Association between ADC and clinicopathological factors (i.e. pathologic T and N stages, tumour location, surgical approach, histologic subtype)	ADC values ≤ 1.5×10−3 mm2/sec were associated with a negative prognosis in the total population (log-relative risk 1.73, standard error 0.56, p = 0.002) and in the surgery-only (log-relative risk 1.97, standard error 0.66, p = 0.003) and chemotherapy (log-relative risk 2.93, standard error 1.41 p = 0.03) groups ADC values ≤ 1.5×10−3 mm2/sec were associated with other significant prognostic factors, especially pathologic T and N stages (p < 0.001 and p < 0.017, respectively)
Liu et al.[61]	Gastric cancer (tubular or papillary adenocarcinoma, poorly cohesive adenocarcinoma, signet-ring cell carcinoma)	CT texture parameters in the arterial (AP) and portal venous phase (PVP)	Diagnostic accuracy and correlation between preoperative CT texture parameters and pathological stage	Maximum frequency in the AP and mean, maximum frequency, mode in the PVP correlated positively with T stage, N stage, and overall stage (all p < 0.05) Entropy in the PVP correlated positively with N stage (p = 0.009) and overall stage (p = 0.032) Skewness in the AP had the highest AUC (0.822) in identifying early from advanced gastric cancers At multivariate analysis four parameters (maximum frequency, skewness, entropy in the PVP, and differentiation degree from biopsy) allowed prediction of lymph node metastasis, distinguishing gastric cancers with and without lymph node metastasis with an AUC of 0.892
Liu et al.[62]	Gastric cancer (tubular adenocarcinoma, papillary adenocarcinoma, poorly cohesive adenocarcinoma, signet-ring cell carcinoma, mucinous carcinoma, mixed types)	CT texture parameters	Prediction of histopathological characteristics	Mean attenuation, maximum attenuation, all percentiles and mode derived from PVP images correlated significantly with differentiation degree and Lauren classification (r = −0.231 ~ −0.324, 0.228 ~ 0.321, respectively; p < 0.05) Standard deviation and entropy derived from AP images correlated significantly with Lauren classification (r = −0.265,–0.222, respectively; p < 0.05) Standard deviation and entropy on AP images were significantly lower in cancers with than without vascular invasion (p = 0.035 and p = 0.031, respectively) Minimum attenuation on AP images was significantly higher in cancers with than those without vascular invasion (p = 0.028)
Liu et al.[63]	Gastric cancer	CT texture parameters	Correlation between CT texture parameters and immunohistochemical markers (E-cadherin, Ki67, VEGFR2 and EGFR)	Standard deviation, width, entropy, entropy, correlation and contrast from AP and PVP were significantly correlated with E-cadherin expression level (all p < 0.05) The skewness from the AP and the mean and autocorrelation from the PVP were negatively correlated with Ki67 expression level (all p < 0.05) Width, entropy and contrast from the PVP were positively correlated with VEGFR2 expression level (all p < 0.05) CT texture analysis had an AUC of 0.612÷0.715 for predicting E-cadherin, Ki67 and VEGFR2 expression levels
Kim et al.[64]	Advanced gastric cancer	CT texture parameters	Prediction of occult peritoneal carcinomatosis (PC) detected at surgery	Patients with occult PC showed significantly higher average, entropy, standard deviation, and lower correlation (p < 0.004 for all) Entropy was a significant independent predictor for occult PC, with a cut-off value >7.141 applied to the validation cohort yielding 80% sensitivity and 90% specificity for prediction of occult PC
Li et al.[65]	Gastric cancer following curative resection	CT radiomics features	Prognosis prediction	The radiomics nomogram incorporating the radiomics signature and significant clinicopathological risk factors (T stage, N stage, and differentiation) exhibited significant prognostic superiority over clinical nomogram and radiomics signature alone (Harrell concordance index of 0.82 vs 0.71 and 0.82 vs 0.74, respectively; p < 0.001) Five radiomics features associated with prognosis were correlated with at least one clinicopathological characteristic, including differentiation, tumor size, N stage, TNM stage, and neural invasion (Spearman’s rho coefficient 0.26–0.38, p < 0.05)
Li et al.[66]	Locally advanced gastric cancer	CT radiomics signature	Outcome prediction of neoadjuvant chemotherapy	One cross-combination machine-learning method derived from AP images had an AUC >0.6 12 cross-combination machine-learning methods derived from PVP images had an AUC >0.6 A feature selection method based on linear discriminant analysis +random forest classifier achieved a significant prognostic performance for PVP images (AUC 0.722 ± 0.108, accuracy 0.793, sensitivity 0.636, specificity 0.889; p = 0.041)
Jiang et al.[67]	Gastric adenocarcinoma	CT radiomics signature derived from PVP images	Prediction of survival and response to chemotherapy	A radiomics signature consisting of 19 selected features was significantly associated with DFS (HR 1.744, p < 0.0001) and OS (HR 3.308, p < 0.0001) and was as independent prognostic factor Incorporating the radiomics signature into the radiomics-based nomograms resulted in better performance for the estimation of DFS and OS than TNM staging and clinicopathological nomograms (p < 0.0001) Stage II and III patients with higher radiomics scores exhibited a favorable response to chemotherapy
Caivano et al.[68]	Gastric adenocarcinoma	Preoperative ADC	Diagnostic accuracy compared to conventional MRI with pathology as gold standard	Gastric cancer tissue had lower ADC values than normal gastric walls (0.811 ± 0.300 vs 1.503 ± 0.430×10−3 mm2/s, p < 0.05) Metastatic lymph nodes had lower ADC values than nonmetastatic lymph nodes (1.70 ± 0.40 vs 2.10 ± 0.22×10−3 mm2/s, p < 0.05) The T factor accuracy of conventional MRI and DWI was 73% and 80%, respectively The N staging accuracy of conventional MRI and DWI was 80% and 93%, respectively
Joo et al.[69]	Gastric cancer	MRI with and without DWI, CT	Diagnostic accuracy for preoperative staging with pathology as gold standard	For N staging, MRI with DWI demonstrated higher sensitivity, but lower specificity (86.7% and 58.8%, respectively) than MRI without DWI (50.0% and 94.1%) or CT (43.3% and 100%) (p < 0.05)
He et al.[70]	Gastric cancer	ADC	Correlation between ADC and HER2 expression	Significant correlation between mean ADC values and HER2 status (r = 0.312, p = 0.037) and scores (r = 0.419, p = 0.004) Mean ADC values of HER2+ cancers were significantly higher than those of HER2- tumours (1.211 vs 0.984 × 10−3 mm2/s, p = 0.020) Minimal ADC values of HER2+ cancers were significantly higher than those of HER2- tumours (1.105 vs 0.905 × 10−3 mm2/s, p = 0.036)
Zongqiong et al.[71]	Gastric cancer (tubular adenocarcinoma, signet-ring cell carcinoma, mucinous adenocarcinoma)	Perfusion CT	Tumour grading	BF was higher in poorly and moderately differentiated cancer than in well differentiated cancer (138.59 ± 38.09 and 110.01 ± 31.90 vs 75.28 ± 6.81 ml/100 g/min, p < 0.05) BV was higher in poorly and moderately differentiated cancer than in well differentiated cancer (21.08 ± 4.11 and 18.18 ± 5.62 vs 9.01 ± 0.94 ml/100g, p < 0.05) PS was higher in poorly and moderately differentiated cancer than in well differentiated cancer (57.50 ± 13.28 and 40.08 ± 15.82 vs 10.05 ± 0.71 ml/100 g/min, p < 0.05)
Zhang et al.[72]	Gastric adenocarcinoma	Perfusion CT	Correlation with histological grading, serosal and lymphatic involvement, distant metastasis, pTNM stage, and microvessel density (MVD) assessed by CD34 immunohistochemical staining	PS was significantly higher in poorly than in moderately differentiated cancers (16.471 ± 9.003 vs 9.558 ± 5.422 ml/100 g/min, p = 0.04) PS was significantly correlated with lymphatic involvement (Spearman’s rho coefficient 0.480, p = 0.038) and pTNM stage (rho = 0.509, p = 0.026)
Liang et al.[73]	Gastric adenocarcinoma	DECT [normalised iodine concentration (NIC) on AP and PVP images]	Correlation with histologic grading, serosal and lymphatic involvement, distant metastasis, pTNM stage and MVD	Both AP and PVP NIC values were significantly higher in poorly differentiated than in moderately differentiated tumours (0.133 ± 0.032 vs 0.102 ± 0.028, p = 0.005 and 0.419 ± 0.066 vs 0.369 ± 0.044, p = 0.013, respectively) Significant correlation between NIC and MVD (AP NIC r = 0.423, p = 0.013; PVP NIC r = 0.542, p = 0.001) Significant difference between the high and low MVD groups with respect to PVP NIC (p = 0.045)
Meng et al.[74]	Gastric mucosal cancer	DECT (IC and NIC on AP and PVP images)	Diagnostic accuracy for differentiating gastric cancer from normal gastric mucosa (NGM) and gastric inflammation (GI)	Mean NIC values of gastric cancer were significantly higher than those of NGM (AP 0.18 ± 0.06 vs 0.12 ± 0.03, PVP 0.62 ± 0.16 vs 0.37 ± 0.08, p < 0.05) Significantly different IC values in gastric cancer compared with GI and NGM (AP 24.19 ± 8.27, 19.07 ± 5.82 and 13.61 ± 2.52 mg ml−1; PVP 28.00 ± 7.01, 24.66 ± 6.5 and 16.94 ± 3.06 mg ml−1, respectively; p < 0.05) NIC and IC in AP had sensitivity of 71.43% and 88.89% in differentiating gastric cancer from NGM NIC and IC in PVP had sensitivity of 88.89% and 90.48% in differentiating gastric cancer from NGM
Pan et al.[75]	Gastric cancer (adenocarcinoma, signet-ring cell carcinoma, mucinous adenocarcinoma)	DECT (keV images, conventional kVp images, NIC)	Diagnostic accuracy with pathology as gold standard	Overall accuracies for T, N and M staging determined with keV images and conventional kVp images were 81.2%, 80.0%, 98.9% and 73.9%, 75.0%, 98.9%, respectively Higher accuracy in N-staging using optimal keV images than conventional kVp images (sensitivity, specificity and accuracy of 90.7%, 67.9% and 82.9% vs 88.8%, 60.7% and 79.2%; p < 0.05) NIC values were significantly different between differentiated and undifferentiated cancer both in AP (0.21 ± 0.08 vs 0.28 ± 0.16, p = 0.02) and PVP (0.54 ± 0.17 vs 0.46 ± 0.12, p = 0.01) NIC values were significantly different between metastatic and nonmetastatic lymph nodes both in AP (0.22 ± 0.09 vs 0.13 ± 0.06, p < 0.001) and PVP (0.47 ± 0.14 vs 0.30 ± 0.12, p < 0.001) A threshold value of 0.145 for AP NIC and 0.333 for PVP NIC yielded a sensitivity and specificity of 84.1%, 67.1% and 89.9%, 67.6%, respectively
Cheng et al.[76]	Early and advanced gastric adenocarcinoma	DECT [IC, NIC, curve slope (λHU) in the PVP and delayed phase (DP)]	Association with Ki-67 protein expression as a marker of cell proliferation	DECT parameters were significantly lower in early than in advanced gastric cancers both in PVP (IC 19.36 ± 2.82 vs 21.25 ± 4.91 mg ml−1; NIC 0.35 ± 0.11 vs 0.42 ± 0.12; λHU 2.20 ± 0.43 vs 2.67 ± 0.63; p < 0.05) and DP (IC 16.89 ± 2.07 vs 19.10 ± 4.07 mg ml−1; NIC 0.43 ± 0.06 vs 0.58 ± 0.14; λHU 1.99 ± 0.35 vs 2.51 ± 0.64; p < 0.01) DECT parameters were positively correlated with Ki-67 grade (Spearman rho 0.818, 0.753, 0.728 in PVP and 0.730, 0.745, 0.468 in DP, respectively; p < 0.01)

Figure 3.

A 49-year-old male with T3N2 advanced gastric cancer without peritoneal seeding showed entropy of 7.05 within the region of interest (a). A 59-year-old female with T3N2 advanced gastric cancer with occult seeding (b) showed entropy of 7.70, higher than the cut-off value (>7.141) obtained from receiver operating characteristic (ROC) curve analysis (c). Reproduced and adapted from.[64]

Among quantitative imaging biomarkers, DWI has shown higher accuracy than conventional MRI for T- and N-staging of gastric adenocarcinoma (80% vs 73% and 93% vs 80%, respectively),[68] and the addition of 3T-MR DWI provides significantly higher sensitivity over conventional MRI or contrast-enhanced CT for the assessment of lymph node metastasis.[69] In a multivariate analysis of 99 patients with gastric cancer (of whom 71 underwent surgery and 28 neoadjuvant chemotherapy, respectively), ADC values of 1.5 × 10−3 mm2/s or lower were associated with a negative prognosis along with pathologic T and N stages, suggesting an important role of DWI in outcome prediction.[58] Overall, such findings seem to reflect a relationship between quantitative imaging biomarkers and different molecular tumour fingerprints that may be exploited to tailor treatment, and to this latter regard it is worth mentioning, for instance, that ADC values were found to correlate with HER2 status.[70]

Perfusion CT and especially DECT can also give a relevant contribution to local staging and prognostic assessment of gastric cancer. Zongqiong et al showed a correlation between perfusion CT parameters (BF, BV and PS values) and the degree of differentiation of gastric adenocarcinoma,[71] whereas Zhang et al reported a significant difference in the PS values between patients with or without lymphatic involvement (p = 0.038), different histological grades (p = 0.04) and TNM staging (p = 0.026).[72] On the other hand, in Liang et al’s study DECT-derived normalised iodine content on arterial and portal venous phase acquisitions was higher in poorly differentiated gastric adenocarcinomas than in moderately differentiated tumours, along with a significant difference in portal venous phase normalised iodine content between tumours with high and low microvessel density, as determined by CD34 staining.[73] Quantitative analysis of DECT imaging parameters for the gastric mucosa can be useful to differentiate malignant from benign gastric mucosal lesions, with iodine concentration and normalised iodine concentration in gastric cancer being significantly different from those in normal and inflammatory mucosa, showing a sensitivity and specificity of about 90% in differentiating gastric cancer from normal mucosa on portal venous phase images.[74] More generally, DECT with monochromatic imaging and quantitative iodine concentration measurements has higher diagnostic accuracy than conventional single-energy CT for T and N staging of gastric cancer (81.2% and 80% vs 73.9% and 75%, respectively), with iodine concentration indexes differing significantly between differentiated and undifferentiated gastric carcinoma, and between metastatic and non-metastatic lymph nodes.[75] The diagnostic potential of DECT may be pushed even further to a molecular level, with quantitative parameters (such as iodine concentration, normalised iodine concentration, and curve slope) showing a significant positive correlation with Ki-67 antigen expression as a marker of tumour proliferation and allowing to differentiate between early and advanced gastric cancer.[76]

Based on the above considerations, the evidence collected so far shows that imaging biomarkers can provide a significant added value to conventional imaging for improving T and N staging of oesophageal and gastric cancers and predicting early response to neoadjuvant CRT and overall prognosis. This might play a pivotal role into identifying the best treatment strategy for patients on a case-by-case basis [e.g. surgery with or without CRT, or neoadjuvant CRT? which degree of surgical radicality? which expected lesion aggressiveness (including conditions with a dramatic impact on treatment choice and prognosis, such as peritoneal dissemination)? any and which potential targets for specific therapies?], possibly leading to better patient outcomes, higher quality of life, and more effective use of healthcare resources.

Conclusions

Imaging biomarkers represent an exciting area of active research that can currently be considered a new frontier of oncologic imaging, owing to their potential to provide a substantial amount of additional information to guide all management steps of cancer patients (from lesion detection and staging to treatment planning, follow-up and prognostic assessment). In particular, the diagnostic workup of upper gastrointestinal cancers should significantly benefit from a combined and thorough analysis of available imaging biomarkers, due to the heterogeneous biological features of such diseases that usually require a complex diagnostic and therapeutic approach. It is likely that in the near future, a growing number of imaging biomarkers and radiomics features will reach the status of validated parameters to be routinely assessed along with established diagnostic imaging criteria, possibly favouring treatment individualisation and improving overall patient outcome.