The emerging field of radiomics in esophageal cancer: current evidence and future potential

Abstract

‘Radiomics’ is the name given to the emerging field of extracting additional information from standard medical images using advanced feature analysis. This innovative form of quantitative image analysis appears to have future potential for clinical practice in patients with esophageal cancer by providing an additional layer of information to the standard imaging assessment. There is a growing body of evidence suggesting that radiomics may provide incremental value for staging, predicting treatment response, and predicting survival in esophageal cancer, for which the current work-up has substantial limitations. This review outlines the available evidence and future potential for the application of radiomics in the management of patients with esophageal cancer. In addition, an overview of the current evidence on the importance of reproducibility of image features and the substantial influence of varying smoothing scales, quantization levels, and segmentation methods is provided.

Introduction

Esophageal cancer continues to affect more than 450.000 people worldwide, making it the eighth most common malignancy and the sixth leading cause of cancer-related mortality (1). Despite recent improvements in staging, multimodality treatment, and peri-operative care, it remains a devastating disease with a 5-year overall survival rate of 15–25% (1–3). Best outcomes are achieved in patients with early carcinoma of the esophagus for which endoscopic mucosal (or submucosal) resection with or without local ablative techniques is now more extensively employed, associated with 5-year survival rates of 60–80% (4,5). Recently, prognosis of locally advanced esophageal cancer has been markedly improved from a 5-year overall survival rate of 23–34% with surgery alone, to 36–47% with the addition of neoadjuvant chemoradiotherapy (nCRT) or peri-operative chemotherapy (6–8). Due to the late onset of symptoms, the majority of patients present at an advanced stage with unresectable or metastatic disease, for which concurrent chemoradiotherapy (CRT) and combination chemotherapy are considered the best palliative options, respectively (9–11).

Current diagnostic work-up consists of endoscopy with biopsy for histopathologic confirmation of the diagnosis and endoscopic ultrasound (EUS) for determination of the local tumor extent (T-stage) and regional lymph node involvement (N-stage) (12,13). Integrated ¹⁸F-fluorodeoxyglucose ( ¹⁸F-FDG) positron emission tomography (PET)/computed tomography (CT) is also used for N-staging, and particularly important for the detection of distant metastasis (M-stage) (14–16). In addition, ¹⁸F-FDG PET/CT is increasingly applied for the detection of interval metastasis after (neoadjuvant) treatment as well as for follow-up after treatment with curative intent (17–20). Unfortunately, these modalities all have their limitations regarding three clinically relevant areas that are in need of improvement, including staging, prediction of response to treatment, and prediction of survival.

It is increasingly recognized that the amount of information currently extracted from available images may be substantially enhanced by quantitative imaging analysis (21). The emerging field of ‘radiomics’ focuses on these improvements of image analysis by extracting large amounts of quantitative image features from volumes of interest on medical images (21). Radiomics approaches can extract more information from medical images by post-processing techniques including quantification of the heterogeneity within a tumor, which is a well-recognized feature of malignancy associated with adverse tumor biology (22). Substantial spatial heterogeneity in metabolism, vasculature, oxygenation, and gene expression is often found in malignant tumors, which relates to chemoradiation-resistance and poor prognosis (23–26). As such, it has been hypothesized that image-based quantification of tumor heterogeneity—through its relation with biologic tumor characteristics may provide important information for staging, predicting response to treatment, and predicting prognosis in cancer patients (21,22,27,28). Indeed, growing evidence suggests that image analysis of tumor heterogeneity could be useful in several cancer types (29–34).

The innovative field of radiomics could provide opportunities in the management of patients with esophageal cancer for improvements in staging, predicting treatment response, and predicting survival. The aim of this review is to outline the current evidence and future potential for the application of radiomics in patients with esophageal cancer.

Texture feature analysis

Both CT and ¹⁸F-FDG PET images are of particular interest when considering quantitative analysis in esophageal tumors, since these modalities are routinely used in clinical practice. CT images are mostly used to extract morphologic information of esophageal tumors, but recent studies suggest that quantitative image features can provide additional information (22,35,36). Most ¹⁸F-FDG PET studies in esophageal cancer quantify metabolic tumor activity solely by using the maximum standardized uptake value (SUVmax) (37,38). However, extracted from a single voxel, SUVmax does not characterize the total activity nor heterogeneity of the ¹⁸F-FDG uptake for the entire tumor (39,40). Recent studies suggest that spatial image information, such as metabolic tumor volume (MTV), total lesion glycolysis (TLG), tumor shape, and texture features, provide more useful information than SUVmax (41–45).

Texture features

Among the studied CT and ¹⁸F-FDG PET quantitative image parameters, texture features are most informative on tumor heterogeneity, and thought to be most closely related to underlying physiologic processes such as vascularization, perfusion, cellular proliferation, and hypoxia (36,46). Texture is defined as a spatial arrangement of voxels allowing extraction of complex image properties (41,46). Different approaches can be used to quantify tumor texture, including model-based fractal analysis and statisticbased methods (47). Model-based fractal analysis methods describe the complexity of an object by identifying the property of self-similarity and roughness of a surface at different levels, and have so far only been described twice in esophageal tumors (48,49). Statistic-based approaches have been most widely used for texture analysis in (esophageal) oncology, and are based on the distribution and spatial relationship of voxel intensity values within an image (47).

Within the statistic-based approaches, first-order statistics represent texture on a global scale calculated from the original voxel intensity values without taking the spatial relationship between voxels into account (e.g., mean, median, percentiles, quartiles, range, interquartile range, standard deviation [SD], coefficient of variation [COV], skewness, kurtosis) (49). Second-order statistics represent texture on a local scale and measure co-occurrence of voxel pairs using grey-level co-occurrence matrices (GLCM; e.g., entropy, energy, homogeneity, contrast/inertia, correlation, dissimilarity) (28). Higher-order statistics capture properties of three or more voxels occurring at specific locations relative to each other, and represent regional texture extracted from grey-level run length matrices (GLRLM; e.g., high/ low grey-level run emphasis, run percentage), grey-level size zone matrices (GLSZM; e.g., high/low intensity zone emphasis, zone percentage), or local texture extracted from neighborhood gray-tone difference matrices (NGTDM; e.g., coarseness, busyness, texture strength, complexity) (28,49,50).

Reproducibility

An overview of studies reporting on the reproducibility and precision of image texture features in esophageal cancer is provided in Table 1. Two studies (in ¹⁸F-FDG PET) have demonstrated that only a limited number of texture features are reproducible with respect to physiological variability as assessed on double baseline scans (46,50). Tixier et al. acquired double baseline ¹⁸F-FDG PET scans within 2–7 days of each other on the same scanner in 16 esophageal cancer patients and reported that the most reproducible features were local entropy, local homogeneity, regional intensity variability, and regional size-zone variability (46). Similarly, in a study from the US including 7 patients who underwent double baseline ¹⁸F-FDG PET scanning within 11–42 days of each other on scanners from different institutions, van Rossum et al. found that the reproducibility was good for (local) second-order and regional higher-order features, but poor for local higher-order features (50). Although the two studies have small samples, these results suggest that only a certain amount of texture features may be used in further research, and other features with poor reproducibility should be abandoned as results will likely not be generalizable. Similar findings were observed by Galavis et al. that found good reproducibility of some texture features but poor reproducibility for others when acquisition modes and reconstruction parameters were varied in ¹⁸F-FDG PET scans of 20 solid tumors (54). To this regard, additional investigation of reproducibility should be encouraged to move this field forward.

Table 1.

Studies on robustness or reproducibility of image features analysis in esophageal cancer

Studies	n	Histology (AC/SCC)	Tumor stage	Treatment	Imaging modality	Imaging timing	Image parameters	Outcome
CT
Ganeshan 2012 (22)	21	14/7	II-IV	NR	Un-enhanced CT	Baseline	Entropy, uniformity 6 smoothing scales	Influence of varying smoothing scales on entropy and uniformity
PET
Tixier 2011 (27)	41	10/31	I-IV	dCRT	18F-FDG PET	Baseline	7 intensity and 31 texture features 4 quantization levels	Influence of varying quantization levels on texture features
Tixier 2012 (46)	16	NR	NR	NR	18F-FDG PET	Double baseline	8 global, 11 regional, and 6 local texture features 5 quantization levels	Test-retest reproducibility of texture features
Hatt 2013 (51)	50	14/36	I-IV	dCRT	18F-FDG PET	Baseline	10 texture features 3 segmentation methods with and without PVC	Influence of varying segmentation methods and PVC on texture features
Dong 2015 (40)	50	0/50	I-IV	Surgery ± adjuvant ChTx/RT or dCRT or ChTx	18F-FDG PET	Baseline	Visual heterogeneity score, coefficient of variation of SUV, and entropy 4 segmentation methods	Influence of tumor heterogeneity on delineated tumor volume using different segmentation methods
Doumou 2015 (48)	64	64/0	NR	NR	18F-FDG PET	Baseline	57 texture features 5 smoothing values 4 segmentation methods 5 quantization levels	Influence of varying smoothing values, segmentation methods, and quantization levels on texture features
Hatt 2015 (52)	112	63/49	I-III	dCRT (39%) or nCRT + surgery (61%)	18F-FDG PET	Baseline	MTV, entropy, dissimilarity, HILAE, and zone percentage 2 calculation methods (for entropy, dissimilarity) quantization levels	Influence of varying calculation methods and quantization levels on correlation between MTV and texture features
van Rossum 2016 (50)	217	217/0	II-III	nCRT + surgery (36% ChTx before nCRT)	18F-FDG PET	Double baseline (in 7 of 217 patients)	69 texture and 12 geometry features 2 baseline scans at different institutions	Test-retest reproducibility of texture and geometry features (in 7 of 217 patients)
Yip 2016 (53)	45	44/1	I-IV	nCRT + surgery	18F-FDG PET	Baseline + after nCRT	MTV, entropy, SRHIE, after nCRT and SZHIE 3 quantization levels 11 registration algorithms for propagated post-treatment contours	Influence of varying quantization levels and propagated post-treatment contours on MTV and texture

¹⁸F-FDG, ¹⁸F-fluorodeoxyglucose; AC, adenocarcinoma; ChTx, chemotherapy; CT, computed tomography; dCRT, definitive chemoradiotherapy; HILAE, high-intensity large-area emphasis; MTV, metabolic tumor volume; nCRT, neoadjuvant chemoradiotherapy; NR, not reported; RT, radiotherapy; PET, positron emission tomography; PVC, partial volume correction; SCC, squamous cell carcinoma; SRHIE, short-run high-intensity emphasis; SUV, standardized uptake value; SZHIE, short-zone high-intensity emphasis.

Influence of smoothing

Besides reproducibility, it is also important that similar measurements from the scan data are obtained when changing parameters such as smoothing, quantization, and segmentation. The ability of texture features to stay similar across variation of these parameters is often referred to as the ‘precision’ of texture features (48). Particularly in CT images, different scales of smoothing (i.e., image filtration) using Laplacian of Gaussian spatial band-pass filters are of importance to be able to reduce image noise and highlight different anatomical spatial scales from fine to medium and coarse texture within the tumor (22,35). Commonly used filter values for smoothing are 1.0 (highlighting fine textures, which may enhance tissue parenchymal features), 1.5–2.0 (highlighting medium textures), and 2.5 (highlighting coarse textures, which may enhance vascular features) (35). Ganeshan et al. showed that CT-based texture features (entropy and uniformity) were influenced by the level of smoothing and significantly associated with tumor stage and survival only after smoothing (22).

Influence of quantization

In both CT and ¹⁸F-FDG PET image post-processing, quantization (i.e., resampling) refers to the important process of resampling the Hounsfield (HU) or SUV levels in the image to a certain number of bins. Choosing the number of bins is a trade-off between gaining texture information accuracy with reduced noise effects (using less bins) and gaining amount of texture information (using more bins). Hence, quantization may influence texture features measurements (48,52). In a recent study with 35 lung cancer patients, Leijennar et al. indeed found that the manner of SUV quantization had a crucial effect on the resulting texture features and their interpretation, emphasizing the importance of standardized methodology in texture analysis (55). The most common quantization method includes the use of a fixed number of discrete bins (e.g., 8, 16, 32, 64, 128 bins) to divide the image SUV range into equally spaced intervals resulting in discretized images with varying bin sizes depending on the SUV range (55). However, this method appeared less appropriate for inter- and intra-patient comparison of texture features in a clinical setting than an alternative method that resamples the SUVs with a fixed bin size in units of SUV (e.g., 0.1, 0.5), maintaining a constant intensity resolution across all images (55).

Tixier et al. showed in 12 esophageal tumors that texture features describing local tumor heterogeneity were insensitive to 5 different quantization values using a fixed number of bins (i.e., 8, 16, 32, 64, or 128 bins), while several regional features were sensitive to the chosen quantization value (46). The same authors described a multi-center series of 555 patients with different types of cancer (including 112 esophageal cancer patients) in which they found that significant texture details are lost when using a quantization of less than 32 bins (52). Also, a higher potential for providing complementary information of texture features (i.e., a lower correlation) with respect to MTV was found using 64 rather than 32 bins (52). Yip et al. reported in a series of 45 patients that the value of texture features for predicting response to nCRT was highest when a quantization level of 128 was chosen (53). Doumou et al. found that 51 of 57 studied texture features showed poor agreement across varying quantization levels with a fixed number of bins (i.e., 8, 16, 32, 64, 128 bins), which stresses the need for further evaluation and standardization of quantization in future studies (48). So far, no studies in esophageal cancer patients reported on the influence of varying quantization using a fixed bin size rather than a fixed number of bins.

Influence of segmentation

Accurate segmentation (i.e., contouring) of the tumor volume is crucial for computing texture features (40,48,51). Many segmentation methods have been proposed including manual delineation, fixed or adaptive thresholding, and multiple (semi-)automatic algorithms, but no consensus seems to emerge (56,57). Most popular segmentation methods in esophageal cancer literature include manual contouring(35,36,40,53,58), thresholding methods capturing aligned voxels with SUV values of ≥2.5 (39,40,58–60) or with values ≥30–60% of the maximum intensity or grey-level (22,40,48,49,51), semi-automatic gradient-based contouring (50), and an automatic fuzzy locally adaptive Bayesian (FLAB) method (41,46,51,52).

Tumor heterogeneity is one of the most important factors influencing the results of different segmentation methods (40). Dong et al. recently demonstrated in 50 patients with esophageal squamous cell carcinoma that in tumors with a large size and high ¹⁸F-FDG uptake heterogeneity, large differences in delineated tumor volume across various manual and thresholding segmentation methods existed (40). More specifically, these authors suggest that in large or highly heterogeneous tumors one must be cautious to use frequently applied and relatively simple threshold-based segmentation methods (40). Similarly, Hatt et al. demonstrated that thresholding methods and automatic FLAB contouring led to substantially different functional volumes, significantly affecting some texture features (e.g., dissimilarity, size-zone variability), while not affecting others (e.g., entropy, homogeneity, zone percentage) (51). Doumou et al. found that varying the relative threshold (45%, 50%, 55%, or 60% of SUVmax) resulted in moderate agreement in second-order (regional) features, but in poor agreement in higher-order (regional and local) features (48). Besides the high sensitivity to tumor heterogeneity, thresholding techniques are also sensitive to motion artifacts, noise and contrast variations, leading to disappointing results for small and non-spherical tumors (56,57). Manual delineation is easy to apply, but time-consuming, susceptible to window-level settings, suffering from intra- and inter-observer variability, and results strongly depend on experience of the reader (56). In general, (semi-)automatic segmentation methods are able to provide superior accuracy, reproducibility, and robustness for tumor volume contouring compared with manual and thresholding methods, and should therefore be preferred (61,62).

Influence of tumor volume

Several texture features have appeared highly correlated with esophageal tumor volume suggesting a certain level of dependency (50,52). When high correlations between parameters exist, an added contribution over each other is unlikely. As such, in the previously mentioned French series of 555 patients with different types of cancer (including 112 esophageal cancer patients), Hatt et al. found that the complementary information of tumor volume and heterogeneity increased substantially with larger tumors (52). In fact, added value of texture features over tumor volume alone for outcome prediction was found in tumors ≥10 cm³ only (52). However, instead of excluding tumors smaller than 10cm3 in future texture studies, they recommended that the correlation of texture features and tumor volume should always be reported to show whether texture and volume provide independent or redundant information (52).

Staging

Accurate tumor staging is crucial for determining prognosis and treatment decision-making in individual patients. EUS is the current modality of choice for primary tumor staging, with reported accuracies for distinguishing T-stages of 53–94% (median, 83%) and better performance in advanced compared to early disease (63,64). Disadvantages of EUS include the invasiveness of the technique, a failure rate of 14–25% due to stenotic tumors preventing passage of the endoscope, and the strong dependence of diagnostic performance on the experience of the endoscopist (65–67). CT is inferior to EUS in the evaluation of T-stage, but CT is useful for predicting surgical resectability by excluding tumors that show ingrowth into surrounding structures (66,68). Regional lymph node involvement is generally evaluated using EUS (sensitivity 80%, specificity 70%), CT (sensitivity 50%, specificity 83%), and ¹⁸F-FDG PET (sensitivity 51%, specificity 84%) (69,70). For the detection of distant metastasis, ¹⁸F-FDG PET provides additional diagnostic information over CT in 5–28% of patients at initial presentation (14). Clearly, current clinical staging is suboptimal and in need of improvement (71).

Radiomics

Three studies (1 using CT and 2 using 18F-FDG PET) reported on the potential value of texture features analysis for staging in esophageal cancer (Table 2) (22,59,60). Ganeshan et al. related two tumor heterogeneity features on unenhanced CT (entropy and uniformity) on 6 different smoothing scales (fine to coarse details) to the clinical American Joint Committee on Cancer (AJCC) stage based on ¹⁸F-FDG PET, CT, and EUS (22). They found that tumor heterogeneity was significantly greater in patients with clinical stage III-IV compared to stage II (22). However, the potential additional value of texture features beyond conventional staging could not be studied as conventional staging was considered the reference standard an no comparison with pathologic tumor stage was performed.

Table 2.

Studies on the value of image features analysis for staging in esophageal cancer

Study	n	Histology (AC/SCC)	Tumor stage	Treatment	Imaging modality	Imaging timing	Image parameters	Outcome
CT
Ganeshan 2012 (22)	21	14/7	II-IV	NR	Un-enhanced CT	Baseline	Entropy, uniformity 6 smoothing scales	Clinical AJCC stage (PET-, CT-, and EUS-based)
PET
Dong 2013 (59)	40	0/40	I-III	Surgery alone	18F-FDG PET	Baseline	SUVmax, entropy, energy	Pathologic AJCC stage, T-stage, and N-stage
Ma 2015 (60)	36	0/36	I-III	Surgery alone	18F-FDG PET and 18F-FLT PET	Baseline	SUVmax, SUVmean, entropy, angular second moment, contrast, correlation, inverse differential moment, tumor length, eccentricity	Pathologic AJCC stage, T-stage, and N-stage

¹⁸F-FDG, ¹⁸F-fluorodeoxyglucose; ¹⁸F-FLT, ¹⁸F-fluorothymidine; AC, adenocarcinoma; AJCC, American Joint Committee on Cancer; CT, computed tomography; EUS, endoscopic ultrasound; NR, not reported; PET, positron emission tomography; SCC, squamous cell carcinoma; SUV, standardized uptake value.

Dong et al. correlated ¹⁸F-FDG PET-based SUVmax, entropy, and energy before surgery with pathologic AJCC-stage, T-stage, and N-stage (59). Most of the studied correlations were weak to moderate only, with the exception of two strong correlations (AJCC-stage and entropy, Spearman’s r=0.63; T-stage and entropy, Spearman’s r=0.69) (59). Similar to the British study reported by Ganeshan et al. (22), higher clinical stage and node-positive tumors were associated with increased tumor heterogeneity (i.e., higher entropy) (59). In ROC curve analysis, an entropy value above 4.7 yielded a sensitivity of 78% and specificity of 73% for predicting pathologic AJCC-stage III as opposed to stage I-II (59). Unfortunately, a multivariable analysis to determine the potential incremental value of the entropy value to predict pathologic stage beyond conventional staging modalities (e.g., EUS, CT) was not performed (59).

Ma et al. included 36 patients who underwent both ¹⁸ F-FDG and ¹⁸F-fluorothymidine (¹⁸F-FLT) PET, and compared the performance of 2 intensity, 2 geometry, and 5 texture features of both modalities for staging with pathologic AJCC-stage, T-stage, and N-stage as reference (60). They found that ¹⁸F-FDG PET features showed more significant associations with pathologic AJCC-stage and TN-stage than ¹⁸F-FLT PET features (60). Interestingly, SUVmax, tumor length, and eccentricity appeared more important than the studied texture features (e.g., entropy, correlation, contrast) for staging (60). Unfortunately, ROC curve analysis and multivariable analysis adjusting for conventional staging modalities were lacking, impairing proper interpretation of potential added value in clinical practice (60).

Prediction of treatment response

Esophageal tumors tend to respond differently to neoadjuvant chemo(radio)therapy or definitive CRT. Adenocarcinomas demonstrate a pathologic complete response (pCR; i.e., complete disappearance of viable tumor cells) to chemotherapy or CRT in 8–9% or 23–28% of patients, respectively (6,72,73), whereas squamous cell carcinomas have a pCR rate of 49% after CRT (6). A pCR is associated with favorable disease-free and overall survival rates, and it has been speculated that accurate identification of pCR prior to surgery could yield an organ-preserving approach avoiding unnecessary surgical morbidity (74–76). On the other hand, it is likely that non-responders to CRT (18–25%) or to chemotherapy (44–58%) are harmed by the toxicity of these therapies without prognostic benefit (6,8,77,78). Early identification of non-responders before or during treatment would be beneficial for this group as ineffective therapy could be modified or discontinued (advancing surgery without detrimental delay in the curative setting) (79).

Several diagnostic strategies have been proposed to predict response to treatment in esophageal cancer. The Response Evaluation Criteria in Solid Tumors (RECIST) method is often used for pre- and post-treatment CT scanning in the evaluation of response, but yields a poor sensitivity (33–55%) and moderate specificity (50–71%) for pathologic response (80). In fact, RECIST did not demonstrate any correlation with treatment response nor prognosis in a recent study in patients with esophageal cancer (81). Post-treatment endoscopic biopsy has a high specificity (91%), but poor sensitivity (35%) for detecting residual cancer, whereas EUS after treatment yields a high sensitivity (96%), but very low specificity (11%) (82). Sequential ¹⁸F-FDG PET-based SUV measurements are able to predict treatment response with a moderate sensitivity (67%) and specificity (68%) (37). In addition, some clinical parameters have been repeatedly found to yield minor—but independent—predictive ability for treatment response (i.e., gender, clinical T-stage, and histologic differentiation grade) (50,76,83). Unfortunately, so far even combinations of modalities and clinical parameters do not yield sufficient predictive ability for pathologic response to guide treatment decision-making in routine clinical practice, and a tool with improved accuracy is highly desired (37,50,84).

Radiomics

An overview of studies reporting on the value of radiomics for the prediction of treatment response in esophageal cancer is presented in Table 3. Studies that performed imaging before and after treatment reported that tumor heterogeneity generally decreased following treatment (36,50). It has been hypothesized that tumors could be rendered more homogeneous following treatment due to a reduction in cellular density and interstitial pressure, and normalization of the vasculature with improved intra-tumor perfusion and oxygenation (36).

Table 3.

Studies on the value of image features analysis for the prediction of treatment response in esophageal cancer

Study	n	Histology (AC/SCC/ other)	Tumor stage	Treatment	Imaging modality	Imaging timing	Image parameters	Outcome
CT
Yip 2015 (36)	31	22/9/0	I-IV	nChTx + Surgery	Contrast-enhanced CT	Baseline + after nChTx	Entropy, uniformity, mean grey-level intensity, kurtosis, skewness, and SD 4 smoothing scales	Pathologic response (TRG* 1–3 vs. 4–5)
PET
Tixier 2011 (27)	41	10/31/0	I-IV	dCRT	18F-FDG PET	Baseline	7 intensity and 31 texture features 4 quantization levels	Clinical response (based on CT; RECIST: CR vs. PR vs. non-R)
Hatt 2013 (51)	50	14/36/0	I-IV	dCRT	18F-FDG PET	Baseline	10 texture features 3 segmentation methods With and without PVC	Clinical response (based on CT; RECIST: CR + PR vs. non-R)
Tan 2013–1 (58)	20	17/3/0	II-III	nCRT + Surgery	18F-FDG PET	Baseline + after nCRT	34 intensity, texture, and geometry features	Pathologic response (TRG* 1–2 vs. 3–5)
Tan 2013–2 (85)	20	NR	NR	nCRT + Surgery	18F-FDG PET	Baseline + after nCRT	SUVmax, SUVpeak, TLG, 8 texture features, and 19 histogram distances	Pathologic response (TRG* 1–2 vs. 3–5)
Zhang 2014 (39)	20	17/3/0	II-III	nCRT + Surgery	18F-FDG PET	Baseline + after nCRT	9 intensity, 8 texture, and 15 geometry features, TLG, and 16 clinical features	Pathologic response (TRG* 1–2 vs. 3–5)
Ypsilantis 2015 (49)	107	86/20/1	II-IV	nChTx + Surgery	18F-FDG PET	Baseline	More than 100 texture features vs. convolutional neural network (3S-CNN) trained directly from scans	Pathologic response (TRG* 1–3 vs. 4–5)
van Rossum 2016 (50)	217	217/0/0	II-III	nCRT + Surgery (36% ChTx before nCRT)	18F-FDG PET	Baseline + after nCRT	69 texture and 12 geometry features 2 baseline scans at different institutions	Pathologic response (TRG† 1 vs. 2–4)
Yip 2016 (53)	45	44/1/0	I-IV	nCRT + Surgery	18F-FDG PET	Baseline + after nCRT	MTV, entropy, SRHIE, SZHIE 3 quantization levels 11 registration algorithms for propagated post-treatment contours	Pathologic response (downstaged vs. upstaged or equal pathologic TN-stage compared to baseline clinical TN-stage)

^*According to Mandard et al. (86)

^†according to Chirieac et al. (75)

¹⁸F-FDG, ¹⁸F-fluorodeoxyglucose; 3S-CNN, three-slices convolutional neural network; AC, adenocarcinoma; ChTx, chemotherapy; CR, complete response; CT, computed tomography; dCRT, definitive chemoradiotherapy; MTV, metabolic tumor volume; nChTx, neoadjuvant chemotherapy; nCRT, neoadjuvant chemoradiotherapy; non-R, non-response; NR, not reported; RECIST, Response Evaluation Criteria in Solid Tumors; PET, positron emission tomography; PR, partial response; SCC, squamous cell carcinoma; SD, standard deviation; SRHIE, short-run high-intensity emphasis; SUV, standardized uptake value; SZHIE, short-zone high-intensity emphasis; TLG, total lesion glycolysis; TRG, tumor regression grade.

Yip et al. studied the value of contrast-enhanced CT image features analysis before and after neoadjuvant chemotherapy in 31 patients for the prediction of good versus poor pathologic response [tumor regression grade (TRG) 1–3 vs. 4–5] (36). Statistical significance was not reached for any of the univariable associations between image features and pathologic response with the exception of pre- and post-treatment SD, which however disappeared after correction for multiple testing (36). They also studied the value of >100 baseline ¹⁸F-FDG PET texture features versus a three-slices convolutional neural network (3S-CNN; which is trained directly from scans rather than ‘manually’ calculated) for the prediction of good versus poor pathologic response (n=217) (49). They found that 3S-CNN outperformed texture features analysis resulting in a sensitivity of 81% and specificity of 82% (49), but this finding has not yet been validated in other studies.

In two studies, a French group determined the associations of tumor texture features on baseline ¹⁸F-FDG PET scans with clinical response to definitive CRT (41,51). In their first study (n=41), the authors reported superior univariable discriminatory ability (area-under-the-curves [AUCs] 0.82–0.89) of several texture features (i.e., homogeneity, entropy, intensity variability, and size-zone variability) over SUVmax and SUVmean (AUCs 0.59–0.70) for the prediction of clinical complete response or nonresponse (41). Similarly, in their second study with a partly overlapping study population (n=50), good univariable discriminatory ability (AUCs 0.80–0.90) was achieved with several image features (i.e., MTV, entropy, homogeneity, dissimilarity, intensity variability, and zone percentage) for the prediction of clinical non-response (51). Important limitations of these studies (41,51) include the suboptimal reference standard defined by the CT-based RECIST method—which is known to correlate poorly to true (pathologic) response and survival (80,81)—and the lack of multivariable prediction modeling (adjusted for clinical parameters and other predictive modalities) impairing proper interpretation of potential incremental value in clinical practice.

Based on baseline and post-treatment ¹⁸F-FDG PET scans, investigators from the US aimed to predict pathologic response (TRG 1–2 vs. 3–5) to nCRT in the same 20 patients with esophageal cancer in three separate articles (39,58,85). By extracting 34 intensity, texture, and geometry features at both time points, they found that changes of features over treatment (Δfeatures) appeared more predictive of response than pre- or post-treatment assessment alone (58). Baseline skewness, ΔSUVmean, post-treatment inertia (contrast), correlation, and cluster prominence were found to be significant predictors of pathologic response in univariable analysis (AUCs 0.76–0.85) (58). In the second study, cross-bin histogram distance features were studied (capturing both ¹⁸F-FDG uptake distribution and longitudinal information), resulting in slightly higher prediction accuracies than texture features (85). This finding requires validation as to date no other studies have reported on cross-bin histogram distance features in esophageal cancer imaging. In the third study, multivariable support vector machine (SVM) and logistic regression models were constructed including 33 ¹⁸F-FDG PET image features as well as 16 clinical parameters (39). SVM models achieved higher accuracy than logistic regression models, particularly in models combining many variables (maximum AUC 1.00 vs. 0.90) (39), but it is important to acknowledge that substantial model overfitting has likely occurred given the small sample size and large amount of predictors included in the modeling, resulting in overoptimistic results.

van Rossum et al. studied the value of clinical parameters along with subjective and quantitative parameters from baseline and post-treatment ¹⁸F-FDG PET scans in 217 patients with esophageal adenocarcinoma for the prediction of pCR as opposed to residual cancer after nCRT (50). Similar to other studies, lower baseline heterogeneity and a greater change towards more homogeneous ¹⁸F-FDG uptake after treatment were associated with better response (50). In multivariable analysis and after internal validation using bootstrapping techniques, both ¹⁸F-FDG PET-based subjective assessment of response and texture features analysis provided incremental value beyond clinical predictors, but this discriminatory improvement did not translate into a clinically relevant benefit as determined by decision-curve analysis (50).

Yip et al. addressed the time-consuming issue of contouring longitudinal scans and investigated the usefulness of 11 different registration algorithms for posttreatment contour propagation in relation to their ability to predict pathologic response (53). They showed that propagated contours could be constructed fast (<30 seconds) and that 3 texture features (e.g., entropy) resulting from most algorithms significantly predicted pathologic responders (AUCs 0.72–0.78), with the exception of fastdemons and fast-free-form deformable algorithms, and rigidly propagated contours, which should therefore not be used (53). An uncommon reference standard was used consisting of pathologic TN-downstaging (responders) versus TN-upstaging or no change in stage (non-responders) as compared to the baseline clinical TN-stage (53). This endpoint was likely suboptimal since it has not been validated as surrogate marker for long-term outcomes and clinical TN-staging is inaccurate in many cases.

Prediction of survival

Accurate stratification of patients according to their expected prognosis is crucial at the time of diagnosis as well as throughout treatment and follow-up. The most important prognostic factor before treatment is the clinical TNM-stage, and -to a lesser extent- the initial SUVmax value (4,87). Endoscopic biopsy, subjective ¹⁸F-FDG PET- based response, and SUVmax after chemo (radio)therapy are other parameters with some prognostic value (38,87,88). After surgery, pathologic TNM-staging, lymph node ratio, extracapsular lymph node involvement, radicality of resection, and pathologic response to neoadjuvant therapy have prognostic impact (75,76,89–92). Despite the availability of these prognostic factors, we are still failing our patients in terms of accurate individualized prediction of survival probability resulting in inaccurate patient selection for different treatment approaches, as for example can be seen from the high number of patients with very early progression (24–41% within 1 year) after treatment with curative intent (93).

Radiomics

Studies on the value of radiomics for the prediction of survival in esophageal cancer are outlined in Table 4. Three studies from the UK have described predictive value for survival using texture features based on baseline unenhanced CT (22) or pre- and post-treatment contrast-enhanced CT (35,36). Ganeshan et al. included 21 patients with esophageal cancer for which the important prognostic factor of treatment was not reported- and studied tumor entropy and uniformity for 6 smoothing scales (22). It was demonstrated that the CT-based coarse uniformity feature was superiorly predictive for overall survival, even resulting in redundancy of clinical TNM-stage and SUVmax in a stepwise forward Cox regression analysis, suggesting substantial overlap in information (22). In the two other studies, tumor entropy, uniformity, mean grey-level intensity, kurtosis, skewness, and SD for 4 quantization levels on baseline and post-treatment contrast-enhanced CT was related to overall survival (35,36). After adjusting for tumor stage and age, post-treatment entropy and uniformity features on a medium to coarse scale remained significant prognostic factors in 36 patients who underwent definitive CRT (35). Yip et al. reported that a relative change in skewness on a fine smoothing scale was associated with survival in 31 patients who underwent neoadjuvant chemotherapy followed by surgery (36).

Table 4.

Studies on the value of image features analysis for the prediction of survival in esophageal cancer

Study	n	Histology (AC/SCC/other)	Tumor stage	Treatment	Imaging modality	Imaging timing	Image parameters	Outcome
CT
Ganeshan 2012 (22)	21	14/7	II-IV	NR	Un-enhanced CT	Baseline	Entropy, uniformity 6 smoothing scales	Overall survival
Yip 2014 (35)	36	9/26/1	I-IV	dCRT (56% ChTx before dCRT)	Contrast-enhanced CT	Baseline + after dCRT	Entropy, uniformity, mean grey-level intensity, kurtosis, skewness, and SD 4 smoothing scales	Overall survival
Yip 2015 (36)	31	22/9/0	I-IV	nChTx + Surgery	Contrast-enhanced CT	Baseline + after nChTx	Entropy, uniformity, mean grey-level intensity, kurtosis, skewness, and SD 4 smoothing scales	Overall survival
PET
Hatt 2015 (52)	112	63/49	I-III	dCRT (39%) or nCRT + Surgery (61%)	18F-FDG PET	Baseline	MTV, entropy, dissimilarity, HILAE, and zone percentage 2 calculation methods (for entropy, dissimilarity) 7 quantization levels	Overall survival

¹⁸F-FDG, ¹⁸F-fluorodeoxyglucose; 3S-CNN, three-slices convolutional neural network; AC, adenocarcinoma; ChTx, chemotherapy; CT, computed tomography; dCRT, definitive chemoradiotherapy; HILAE, high-intensity large-area emphasis; MTV, metabolic tumor volume; nChTx, neoadjuvant chemotherapy; nCRT, neoadjuvant chemoradiotherapy; NR, not reported; PET, positron emission tomography; SCC, squamous cell carcinoma; SD, standard deviation.

In a French study that included baseline ¹⁸F-FDG PET scans of 555 cancer patients (of whom 112 had esophageal cancer and underwent definitive CRT or nCRT followed by surgery), Hatt et al. assessed the value of MTV, entropy, dissimilarity, high-intensity large-area emphasis, and zone percentage for the prediction of overall survival (52). Although MTV and heterogeneity (along with tumor stage) were independent prognostic factors in non-small cell lung cancer, these parameters had less complementary value in esophageal cancer which was attributed to smaller overall volumes (52). The local dissimilarity parameter appeared most predictive for overall survival in the patients with esophageal cancer (52).

Conclusions

Since the first publication on image texture feature analysis in esophageal cancer in the year 2011 (41), the body of evidence on radiomics in this setting has been growing steadily suggesting potential incremental value for staging, prediction of response to chemo(radio)therapy, and predicting survival. As such, radiomics approaches may contribute to the ongoing movement towards more individualized treatment strategies for these patients. An advantage of this emerging field is that it can fit in within existing practice without imposing additional burden to patients, as it involves post-processing techniques on standard CT or ¹⁸F-FDG PET images which are performed as part of routine clinical practice. However, current evidence is still exploratory in nature and further validation in larger studies is required before implementation in clinical practice could be considered.

Acknowledgement and further evaluation of limitations with respect to reproducibility of image features and the substantial influence of varying smoothing scales, quantization levels, and contouring methods is of crucial importance to move this field forward. To this regard, parameters such as local entropy derived from GLCMs (and to a lesser extent uniformity, dissimilarity, or zone percentage) for tumor heterogeneity characterization should be preferred, as these appear most reproducible and robust, and have repeatedly shown high predictive ability for staging, prediction of response, and prediction of survival. Standardization of imaging and radiomics approaches, multivariable prediction modeling focusing on incremental value of radiomics beyond conventional diagnostics and predictors, and validation of findings are key to successful future introduction of radiomics in the clinical management of esophageal cancer.