Artificial intelligence in quantitative imaging of esophageal cancer: A review on radiomics, sarcopenia, and survival modeling

Abstract

Esophageal cancer remains one of the most lethal malignancies worldwide, with survival outcomes varying widely even among patients with similar clinical stages. Recent advances in artificial intelligence (AI) have enabled the extraction of quantitative imaging features, known as radiomics, from routine computed tomography and positron emission tomography/computed tomography scans, offering new opportunities for precision prognostication. At the same time, body composition metrics such as sarcopenia and visceral adiposity have emerged as important predictors of treatment tolerance and overall survival. This article summarizes current evidence on artificial intelligence-based approaches that integrate tumor radiomics and host body composition for survival modeling in esophageal cancer. It outlines methodological frameworks, model performance, and key predictors identified across studies, and discusses challenges related to data harmonization, feature reproducibility, and clinical translation. The combined use of radiomics and body composition analysis through machine learning offers a promising path toward individualized, image-based survival prediction beyond conventional staging systems.

Key Words: Esophageal cancer; Artificial intelligence; Radiomics; Deep learning; Sarcopenia; Survival prediction; Quantitative imaging

Core Tip: This article highlights how artificial intelligence (AI) enables a shift from anatomy-based staging to quantitative, image-driven prognostication in esophageal cancer. By integrating tumor radiomics with AI-derived body composition markers such as sarcopenia, survival models can capture both tumor aggressiveness and host vulnerability from routine imaging modalities. These multimodal approaches consistently outperform conventional staging in survival prediction and risk stratification. Despite challenges in standardization and validation, AI-based quantitative imaging offers a clinically scalable pathway toward personalized survival modeling and precision treatment planning in esophageal cancer.

INTRODUCTION

Esophageal cancer remains a major oncologic challenge, marked by aggressive behavior and wide variability in treatment response and long-term outcomes[1,2]. This continues to impose a severe global health burden, with mortality that remains disproportionately high relative to its incidence. In 2022, it accounted for over half a million new diagnoses worldwide, yet contributed to nearly 450000 deaths, placing it among the leading causes of cancer-related mortality despite being outside the top ten by frequency[3]. This striking imbalance reflects its aggressive course, frequent late presentation, and wide variability in treatment response even with modern multimodality care. A key limitation is that routine evaluation still depends heavily on subjective visual assessment and basic measurements on imaging, which cannot fully represent tumor heterogeneity, subtle biological differences, or early treatment-related changes[4]. These gaps highlight the clinical need for objective, reproducible, and scalable quantitative imaging biomarkers that can strengthen risk stratification, guide treatment intensity, and improve individualized prognostication in real-world practice.

Radiomics has emerged as a promising approach by extracting high-dimensional quantitative features from standard computed tomography (CT), positron emission tomography (PET)/CT, and magnetic resonance imaging (MRI), transforming routine scans into mineable data that may serve as imaging biomarkers for precision oncology[5,6]. Recent systematic reviews suggest that artificial intelligence (AI)-based radiomics can support diagnosis, treatment response prediction, and survival estimation in esophageal cancer; however, evidence remains heterogeneous, and robust external validation is frequently lacking. Similar conclusions have been reported in other contemporary reviews, highlighting that methodological innovation has outpaced real-world clinical adoption[7]. Other than tumor-focused radiomics, increasing attention has been directed toward AI-driven body composition analysis, as outcomes in esophageal cancer are shaped not only by tumor biology but also by host-related factors such as nutritional status, frailty, and physiological reserve[8,9]. Quantitative assessment of sarcopenia and skeletal muscle quality from routine staging CT scans has emerged as a clinically relevant prognostic marker, particularly in a disease characterized by weight loss and treatment-related physiological stress. Advances in automated and semi-automated AI segmentation have improved the reproducibility and feasibility of body composition analysis, enabling its integration into routine workflows[10]. When combined with radiomics and clinical variables, these host-derived imaging biomarkers can be incorporated into machine learning-based survival models to provide more accurate and individualized prognostic estimates for improving risk stratification, guiding treatment intensity, and supporting shared decision-making in esophageal cancer care[11,12]. This article presents an integrated evaluation of tumor radiomics, AI-enabled body composition imaging, and survival modelling, highlighting their collective role in advancing quantitative imaging for esophageal cancer. Relevant literature for this was identified through a structured search of the PubMed database covering publications from January 2015 to January 2026. The search was performed using Boolean combinations of keywords including “esophageal cancer”, “radiomics”, “artificial intelligence”, “deep learning”, “sarcopenia”, “body composition”, and “survival prediction”. Example search strings included “esophageal cancer AND radiomics”, “esophageal cancer AND artificial intelligence AND survival”, and “esophageal cancer AND sarcopenia AND CT imaging”. Articles were screened based on relevance to quantitative imaging biomarkers, AI-driven radiomics, body composition analysis, and survival modeling in esophageal cancer. The included studies reflect a selective but representative overview of the current literature.

ROLE OF RADIOMICS IN ESOPHAGEAL CANCER

Radiomics refers to the systematic extraction of quantitative imaging features that capture spatial heterogeneity, texture patterns, shape characteristics, and intensity distributions in a region or volume of interest[13-15]. These parameters reflect the underlying biological phenomena such as hypoxia, necrosis, vascular heterogeneity, and cellular density. Radiomic features can be broadly categorized into several classes based on the type of information they encode. They are first-order, second-order or texture, shape-based, high order, and deep features (Table 1).

Table 1 Quantitative radiomic features used in esophageal cancer.

Feature category	Primary parameters	Description	Clinical significance
First-order	Mean, median, skewness, kurtosis, entropy, energy	Statistical distribution of voxel intensities without spatial context	Reflects global tumor density, degree of necrosis, and metabolic activity
Second-order (texture)	GLCM, GLSZM, GLRLM, NGTDM	Spatial relationships and patterns between neighboring pixels	Quantifies intratumoral heterogeneity; predicts treatment resistance and aggressive phenotypes
Shape-based	Sphericity, surface-to-volume ratio, compactness, elongation	Three-dimensional geometric and morphological properties	Correlates with longitudinal infiltration and circumferential involvement for T-staging
Higher-order	Wavelet, LoG, square root	Mathematical filters applied to original images to enhance specific frequencies	Captures fine structural details and patterns linked to microvascular complexity and hypoxia
Deep features	CNN latent layers	Abstract patterns learned automatically by neural networks	High predictive power for survival, though often lacking direct biological interpretability

GLCM: Gray-level co-occurrence matrix; GLSZM: Gray-level size zone matrix; GLRLM: Gray-level run length matrix; NGTDM: Neighborhood gray-tone difference matrix; LoG: Laplacian-of-gaussian; CNN: Convolutional neural network.

The radiomics workflow involves image acquisition, tumor segmentation, feature extraction, feature selection, model construction, and validation. Each step introduces potential sources of variability that influence reproducibility and clinical applicability. In esophageal cancer, tumor segmentation is particularly challenging due to motion artifacts, indistinct tumor boundaries, and mucosal disease spread, leading to inter-observer variability that directly propagates into feature instability[16,17]. Automated and semi-automated segmentation methods based on deep learning have shown promise in reducing this variability and enabling scalable feature extraction. Feature selection is critical to avoid overfitting, as radiomic datasets are high-dimensional relative to sample size; therefore, reproducibility filtering, correlation analysis, and penalized modelling approaches are commonly employed[18]. To ensure the clinical validity of the radiomic metrics, recent studies emphasize adherence to the Image Biomarker Standardization Initiative (IBSI), a global effort to standardize the mathematical definitions of radiomic features across different software platforms and imaging kernels.

The extracted radiomic features have served a broad range of purposes in esophageal cancer, extending the role of medical imaging from descriptive or visual-only interpretation to objective, quantitative clinical decision support. In esophageal cancer, radiomics has been applied across the entire disease, including tumor detection and characterization, histological subtype discrimination, staging, assessment of treatment response, prediction of recurrence, and survival prognostication[7,19]. Quantitative radiomic analysis has demonstrated the ability to differentiate malignant from benign esophageal lesions and to assist in distinguishing esophageal squamous cell carcinoma from esophageal adenocarcinoma by capturing differences in voxel intensity distributions, spatial heterogeneity, and longitudinal growth patterns that are not appreciable on routine visual assessment[20]. Within the staging workflow of esophageal cancer, radiomic models have consistently outperformed conventional size-based imaging criteria for predicting lymph node metastasis and local tumor extent, providing a noninvasive means to refine preoperative risk stratification[21,22].

The prognostic value of radiomics in preoperative staging of esophageal cancer is particularly evident in recent efforts to identify occult locally advanced disease that may be missed by standard radiological interpretation. In a pivotal study by Guo et al[7], a high-dimensional CT-based radiomic framework incorporating 1067 features was developed specifically for esophageal cancer to detect sub-visual invasion patterns, achieving an area under the receiver operating characteristic curve of 0.85. This capability is clinically critical in esophageal cancer, as accurate identification of occult locally advanced disease ensures that patients are appropriately selected for neoadjuvant chemoradiotherapy rather than proceeding directly to surgery. Such stratification reduces the risk of incomplete resection and contributes to improved long-term survival outcomes in esophageal cancer[7].

In the setting of neoadjuvant therapy for locally advanced esophageal cancer, radiomics has been most extensively investigated for predicting pathological complete response and tumor regression grade using pretreatment and early on-treatment imaging. Predicting pathological complete response following neoadjuvant chemoradiotherapy remains one of the most important clinical objectives in esophageal cancer management, as it identifies patients who may be candidates for organ-preserving “watch-and-wait” strategies. Conventional response assessment using the Response Evaluation Criteria in Solid Tumors, which relies primarily on macroscopic changes in tumor size, is often inadequate in esophageal cancer because significant cellular destruction and stromal remodeling can occur without substantial reduction in tumor diameter[23]. Consequently, research in esophageal cancer has increasingly focused on multimodal radiomic models that integrate CT, fluorodeoxyglucose positron emission tomography combined with CT, and magnetic resonance imaging. A systematic review and meta-analysis by Kao and Hsu[24] evaluating radiomics for response prediction in esophageal cancer demonstrated that second-order texture features derived from positron emission tomography, particularly gray-level size-zone matrix-based measures of intensity variability, were among the most stable and accurate predictors of pathological complete response, with pooled area under the curve values consistently exceeding 0.81. The development of delta-radiomics, which assesses temporal changes in radiomic features during treatment, has further improved response prediction in esophageal cancer. Several studies have shown that early reductions in entropy and shifts in apparent diffusion coefficient values during the initial weeks of chemoradiotherapy are more predictive of pathological complete response in esophageal cancer than baseline imaging features alone[25,26].

Magnetic resonance imaging-based radiomics has shown particular promise in esophageal cancer owing to its superior soft-tissue contrast and sensitivity to microstructural changes. Diffusion-weighted magnetic resonance imaging and dynamic contrast-enhanced magnetic resonance imaging-derived radiomic features have demonstrated strong associations with treatment response in esophageal cancer. A recent study by Yang et al[27] identified a panel of ten magnetic resonance imaging-derived radiomic features that functioned as noninvasive biomarkers of pathological response in esophageal cancer, suggesting that magnetic resonance imaging-based radiomics may complement or partially replace endoscopic ultrasound for intratreatment monitoring. Similarly, Hirata et al[28] proposed a machine learning framework based on positron emission tomography imaging in esophageal cancer, demonstrating that metabolic tumor volume and total lesion glycolysis, when combined with measures of textural complexity, provide a comprehensive imaging phenotype associated with resistance to neoadjuvant therapy.

Beyond response assessment, radiomics has demonstrated substantial value in predicting disease recurrence and long-term survival in esophageal cancer. Although tumor-node-metastasis (TNM) staging remains the cornerstone of prognostic assessment in esophageal cancer, patients with identical stages frequently exhibit markedly different clinical outcomes. Radiomic signatures have shown the ability to bridge this prognostic gap by capturing biological aggressiveness that is not reflected in conventional staging. CT-based deep learning radiomic nomograms developed for preoperative lymph node assessment in esophageal cancer have not only outperformed clinical nodal staging but have also demonstrated significant associations with three-year overall survival. In a comprehensive analysis by Wesdorp et al[20], esophageal tumors characterized by higher baseline textural entropy and “busyness” were associated with more aggressive clinical behavior and a greater propensity for distant metastasis, even after adequate local control. Integration of radiomic signatures into clinicopathological nomograms for esophageal cancer has yielded net reclassification improvements approaching 27 percent, enabling more personalized postoperative surveillance strategies and improved identification of patients who may benefit from adjuvant systemic therapy[20]. The prognostic utility of radiomics in esophageal cancer is further strengthened when combined with host-related imaging biomarkers such as sarcopenia, allowing simultaneous assessment of tumor aggressiveness and patient physiological reserve.

An emerging frontier in esophageal cancer radiomics is radiogenomics, which seeks to link imaging phenotypes with underlying molecular and genetic characteristics. Several high-impact studies have demonstrated associations between radiomic features and molecular pathways implicated in esophageal cancer progression and treatment resistance, including human epidermal growth factor receptor 2 expression, CD44, and Sonic Hedgehog signaling. For example, skewness in apparent diffusion coefficient distributions on magnetic resonance imaging has been hypothesized to reflect tumor microenvironments in esophageal cancer with reduced stromal content and higher proliferative activity, thereby conferring increased sensitivity to chemoradiotherapy. These findings support the concept that radiomics may serve as a noninvasive surrogate for molecular profiling in esophageal cancer, particularly in tumors with marked spatial heterogeneity or where tissue sampling is limited[29-31].

In addition to CT-based radiomics, PET/CT and MRI radiomics have emerged as important modalities for quantitative imaging in esophageal cancer. PET/CT-based radiomics has received considerable attention because fluorodeoxyglucose positron emission tomography provides functional and metabolic information about tumor activity that cannot be captured by anatomical imaging alone. Radiomic features derived from PET images, including metabolic tumor volume, total lesion glycolysis, and texture-based measures of metabolic heterogeneity, have demonstrated value in predicting treatment response, lymph node metastasis, and survival outcomes. Similarly, MRI-based radiomics is increasingly being investigated due to its superior soft-tissue contrast and sensitivity to microstructural and physiological changes within the tumor. Techniques such as diffusion-weighted imaging and dynamic contrast-enhanced MRI enable extraction of radiomic features reflecting tumor cellularity, perfusion, and microenvironmental heterogeneity. Recent studies suggest that multi-modality models integrating CT, PET/CT, and MRI radiomics can capture complementary structural, metabolic, and functional tumor characteristics, consistently demonstrating improved predictive performance compared with single-modality approaches.

Despite these advances, clinical translation of radiomics in esophageal cancer remains constrained by reproducibility challenges. Variations in image acquisition parameters, including slice thickness, reconstruction kernels, contrast protocols, and scanner manufacturers, can substantially influence feature values and limit generalizability across institutions[32,33]. Moreover, the predominance of single-center esophageal cancer studies and heterogeneity in segmentation and analytical workflows continue to hinder widespread adoption. Many radiomics studies in esophageal cancer are also limited by relatively small sample sizes and retrospective single-center designs, which increase the risk of model overfitting and reduce generalizability. Inconsistencies in data partitioning strategies and the absence of independent external validation cohorts remain common methodological limitations. Increasing adherence to standardized radiomics reporting frameworks such as the IBSI is therefore essential to improve reproducibility and facilitate clinical translation. Addressing these limitations through standardized acquisition protocols, robust external validation, and prospective multicenter studies is essential for radiomics to become a reliable component of routine esophageal cancer care.

AI-DRIVEN SARCOPENIA AND QUANTITATIVE BODY COMPOSITION IMAGING IN ESOPHAGEAL CANCER

Sarcopenia has emerged as a central host-related determinant of prognosis in esophageal cancer, reflecting the profound systemic effects of tumor burden, dysphagia-related malnutrition, and treatment-induced metabolic stress[34,35]. Unlike many other solid malignancies, esophageal cancer frequently presents with early skeletal muscle depletion, even before initiation of neoadjuvant therapy, due to impaired oral intake and cancer-associated cachexia. As a result, survival outcomes in esophageal cancer are increasingly recognized to be shaped not only by tumor biology but also by the patient’s baseline physiological reserve and its dynamic evolution during treatment. AI-based quantitative imaging has enabled objective, reproducible, and scalable assessment of skeletal muscle mass and quality from routine CT scans, positioning sarcopenia as a key imaging-derived biomarker for outcome prediction[36]. Traditional sarcopenia assessment relies on manual measurement of skeletal muscle area at predefined anatomical landmarks, most commonly the third lumbar vertebral level. However, such approaches are time-consuming, subject to inter-observer variability, and limited to muscle quantity alone. Recent advances in AI have transformed body composition analysis by enabling automated segmentation of multiple muscle groups and extraction of higher-order radiomic features that characterize muscle quality and heterogeneity[37,38]. This shift from simple morphometric indices to texture-informed muscle phenotyping has opened new avenues for understanding the prognostic relevance of skeletal muscle beyond cross-sectional area.

A pivotal contribution in this area is the study by Vogele et al[39], which investigated sarcopenia and CT-based radiomics in patients with esophageal and gastric cancer. In a longitudinal cohort of 83 patients undergoing contrast-enhanced CT between 2015 and 2019, sarcopenia was defined using the psoas muscle index and evaluated at diagnosis, after neoadjuvant chemotherapy, and one year after surgery or chemotherapy, demonstrating a significant progressive decline over time and reinforcing sarcopenia as a dynamic process rather than a static baseline condition. Although psoas muscle index-defined sarcopenia alone was not significantly associated with progressive disease, the study provided important methodological insight by applying AI-based radiomic analysis to skeletal muscle. Radiomic features encompassing shape, first-order intensity, and higher-order texture were extracted from the psoas major, erector spinae, and quadratus lumborum muscles, and machine learning classifiers were trained to predict sarcopenia and tumor progression. Among these, the random forest model achieved excellent performance for sarcopenia detection at baseline, with an accuracy of 0.93 and an area under the receiver operating characteristic curve of 0.97, highlighting the added value of radiomics in capturing qualitative muscle characteristics beyond conventional morphometric assessment[39].

Similarly, Liu et al[8] advanced the field by integrating AI-derived body organ analysis and radiomics with survival modelling in a larger cohort of 212 patients with esophageal cancer treated with concurrent chemoradiotherapy. In contrast to traditional lumbar-level approaches, this study focused on skeletal muscle and adipose tissue quantified at the twelfth thoracic vertebral level, addressing a common clinical limitation when abdominal imaging is unavailable. Automated extraction of skeletal muscle and visceral adipose tissue indices was combined with radiomic features derived from both pretreatment and follow-up CT scans. Using a structured modelling pipeline incorporating least absolute shrinkage and selection operator-based feature selection followed by Cox regression, the authors demonstrated that models integrating both baseline and follow-up imaging features significantly outperformed models based on pretreatment data alone. The best-performing model, which combined clinical variables with radiomics and body composition metrics from both time points, achieved an area under the time-dependent receiver operating characteristic curve of 0.91 for two-year overall survival prediction. Importantly, follow-up imaging features, particularly radiomics derived from visceral adipose tissue, emerged as dominant predictors, reinforcing the importance of longitudinal imaging in capturing treatment-induced physiological changes. This study represents a major step toward clinically feasible AI-driven survival prediction by demonstrating that automated T12-level imaging biomarkers can provide robust prognostic information in esophageal cancer.

The broader clinical significance of integrating radiomics and sarcopenia is further contextualized by the systematic review conducted by Peng et al[40], which examined studies combining quantitative imaging biomarkers and body composition metrics for outcome prediction in esophageal and gastroesophageal cancers. This article highlighted the persistently poor prognosis of esophageal cancer, with global five-year survival rates remaining below 25% despite advances in surgery and chemoradiotherapy. The authors emphasized that traditional TNM staging fails to explain the marked heterogeneity in outcomes observed among patients with similar anatomical disease burden, thus highlighting the need for more precise prognostic tools. Across the six studies reviewed, both radiomic features and sarcopenia independently demonstrated prognostic value, while integrated models consistently outperformed clinical staging alone. Skeletal muscle wasting during neoadjuvant therapy emerged as a particularly strong predictor of inferior recurrence-free and overall survival, highlighting sarcopenia as a dynamic and potentially modifiable risk factor. Importantly, the reviewed studies illustrated that combining tumor-derived radiomics with host-derived body composition metrics enables a more holistic representation of disease aggressiveness and patient resilience. However, current studies integrating sarcopenia and radiomics often involve modest patient cohorts and heterogeneous imaging protocols, which may influence the stability of extracted features. The absence of standardized segmentation pipelines and limited external validation across independent datasets also highlights the need for more robust multicenter investigations.

Collectively, current evidence supports AI-driven sarcopenia assessment as a powerful complement to radiomics-based tumor analysis in esophageal cancer. Quantitative imaging of skeletal muscle captures host vulnerability that is invisible to conventional staging and provides critical prognostic information, particularly when assessed longitudinally. When integrated within machine learning-based survival models, sarcopenia and radiomic features jointly enable refined risk stratification, improved survival prediction, and the potential for personalized treatment adaptation[41,42].

SURVIVAL MODELING IN ESOPHAGEAL CANCER

Survival modelling refers to the quantitative prediction of time-to-event outcomes, such as overall survival or disease-specific mortality, by integrating multiple prognostic variables within statistical or machine learning frameworks[43-45]. In esophageal cancer, survival modelling has particular clinical importance because patients with identical TNM stages often experience markedly different outcomes. Traditional prognostic tools rely primarily on anatomical staging and a limited number of clinicopathological variables, which are insufficient to capture tumor aggressiveness, treatment response heterogeneity, and host-related vulnerability from a comprehensive perspective. AI-based survival modelling seeks to overcome these limitations by learning complex, non-linear relationships from large, multidimensional datasets and translating them into individualized risk estimates that can guide clinical decision-making[46,47].

A contribution in this domain is the large-scale study by Zhang et al[48], which developed machine learning-based survival prediction models specifically for esophageal squamous cell carcinoma. Using a training cohort of 1954 patients and an independent validation cohort of 487 patients, the authors evaluated six machine learning approaches, including random forest, elastic net, gradient boosting machine, and a machine learning-extended Cox proportional hazards model. The study demonstrated that AI-enhanced survival models significantly outperformed the American Joint Committee on Cancer eighth edition staging system in discriminative ability. Importantly, the extended Cox proportional hazards, elastic net, and random forest models showed comparable and robust performance, highlighting that classical survival frameworks, when augmented by machine learning-based feature selection, remain highly effective and interpretable. Risk scores derived from these models stratified patients into low-, intermediate-, and high-risk groups with clearly separated three-year overall survival probabilities, confirming their clinical relevance[48]. Building on the need for interpretability, Tuersun et al[49] further advanced AI-based survival modelling by developing explainable machine learning models with external validation in a Chinese cohort. By comparing Cox proportional hazards models with advanced algorithms such as random survival forest, DeepSurv, and neural multi-task logistic regression, the authors demonstrated that neural multi-task logistic regression achieved the best overall predictive performance across one-, three-, and five-year survival horizons. Crucially, this study incorporated formal interpretability analyses, including time-dependent variable importance and survival SHapley additive interpretation, confirming that established clinical factors such as metastatic stage, nodal involvement, age, and treatment modalities consistently drove predictions. This work addressed a key barrier to clinical adoption by showing that high-performing AI survival models can remain transparent and clinically interpretable.

While survival modelling focuses on outcome prediction after diagnosis, its clinical impact is closely linked to early detection and risk stratification. This connection is emphasized in comprehensive reviews by Zhang et al[50] and Yan et al[51], which place AI-based survival prediction within the broader framework of esophageal squamous cell carcinoma management. These authors highlight that survival outcomes are strongly influenced by disease stage at diagnosis. Machine learning models integrating epidemiological risk factors, minimally invasive cytological sampling, and imaging data have demonstrated high accuracy in identifying high-risk individuals, enabling targeted endoscopic screening while avoiding unnecessary procedures in low-risk populations. Moreover, AI-assisted endoscopic systems have shown expert-level performance in detecting early cancer, assessing invasion depth, and reducing missed lesion rates in real-world clinical settings. These diagnostic advances directly influence survival modelling by shifting patients toward earlier stages at diagnosis, where predicted survival trajectories are substantially more favorable. AI-based survival models therefore do not operate in isolation but are embedded within a comprehensive clinical picture that includes right from screening, diagnosis, treatment selection, to longitudinal outcome prediction (Figure 1). Although machine learning-based survival models demonstrate promising predictive performance, methodological concerns remain. Many published models are developed using relatively small datasets with complex feature spaces, creating a substantial risk of overfitting if appropriate cross-validation and independent validation cohorts are not employed. Future studies should prioritize larger multicenter datasets, transparent reporting of data partitioning strategies, and adherence to standardized radiomics frameworks to ensure reproducibility.

Figure 1 Conceptual workflow for artificial intelligence-based survival modeling in esophageal cancer. AI: Artificial intelligence; TNM: Tumor-node-metastasis.

Several other AI applications are also rapidly advancing in esophageal cancer management. AI-assisted endoscopic imaging represents one of the most clinically mature applications, with multiple studies and randomized trials demonstrating improved detection of early esophageal neoplasia, enhanced characterization of dysplastic lesions, and reduced missed lesion rates during screening endoscopy. AI models are also being developed for prediction of lymph node metastasis and preoperative staging, which directly influence treatment selection, surgical planning, and the decision between primary surgery and neoadjuvant therapy. In addition, emerging research is exploring AI-based prediction of radiotherapy toxicity, which may help personalize radiation dose planning and improve treatment tolerance in patients undergoing chemoradiotherapy. Collectively, these developments highlight the growing importance of multimodal AI integration across imaging, endoscopy, and clinical data, which is expected to play an increasingly important role in improving diagnosis, risk stratification, and personalized management of esophageal cancer.

TECHNICAL CHALLENGES IN AI-DRIVEN QUANTIFICATION IN ESOPHAGEAL CANCER

While the integration of AI-driven radiomics and body composition analysis offers a transformative outlook for esophageal cancer care, several hurdles must be cleared before these tools become standard in the oncology clinic. One of the most significant barriers is the lack of standardization across different medical centers, as radiomic features are highly sensitive to technical noise like variations in CT scanner manufacturers, slice thickness, and contrast enhancement protocols that can alter feature values even in the same patient. Several methodological studies have demonstrated that more than 50% of extracted radiomic features may vary significantly when imaging protocols or scanners differ, highlighting the magnitude of the reproducibility challenge in radiomics research. Evaluations of the radiomics literature have also shown that only about 30% of published studies report sufficient technical details regarding imaging acquisition and preprocessing steps, which further limits independent validation and reproducibility. This variability has important implications for clinical translation, as radiomics models developed using inconsistent imaging protocols may fail to perform reliably when applied to external datasets or routine clinical practice, thereby limiting their adoption as decision-support tools for diagnosis and treatment planning. For AI models to be reliable and reproducible, they must be validated on diverse, multi-institutional datasets to ensure that a “high-risk” signature in one hospital remains consistent in another, making adherence to the IBSI guidelines essential. Furthermore, the first step in quantification, known as segmentation, remains an “ill-posed” problem in the esophagus due to its hollow, contractile nature and poor contrast with surrounding tissues like the aorta or heart. AI algorithms frequently struggle with “segmentation drift” or motion artifacts from cardiac pulsation that blur the edges of the tumor, introducing errors into shape-based measurements. This technical instability is compounded by the “reconstruction effect”, where different kernels can significantly alter texture features like entropy, making a tumor appear more heterogeneous simply due to the scanner's settings rather than the tumor’s biology. To achieve true quantification, researchers must employ techniques like ComBat harmonization to remove the signature of the machine so that only the biological signature of the cancer remains. Similarly, quantifying muscle quality or myosteatosis is complex because small variations in contrast-enhancement timing can change the Hounsfield Unit values of the muscle, potentially leading to an incorrect diagnosis of sarcopenia. Another important methodological challenge is the curse of dimensionality, where a very large number of radiomic features are extracted from relatively small patient cohorts. This reflects the well-known “small-n-large-p” problem in radiomics research, in which hundreds or even thousands of imaging features are analyzed using limited sample sizes. When the number of candidate features substantially exceeds the number of patients, predictive models become highly susceptible to overfitting, resulting in overly optimistic performance estimates that may fail to generalize to independent datasets. In the context of survival modelling, appropriate statistical design is therefore essential, including maintaining adequate events-per-variable ratios, which are commonly recommended to be at least 10-20 outcome events for each predictor variable included in the model. Addressing this challenge requires careful feature selection, adequate sample sizes, and validation of models using independent cohorts to ensure robustness and generalizability. A significant challenge in the clinical adoption of AI models is the “black box” nature of many machine learning and deep learning algorithms. In clinical oncology, predictive models must not only demonstrate high accuracy but also provide interpretable and transparent reasoning to support clinical decision-making. Without interpretability, clinicians may be reluctant to rely on algorithmic predictions, particularly when treatment decisions involve high-risk interventions such as surgery, chemoradiotherapy, or treatment intensification. Consequently, increasing attention has been directed toward explainable AI approaches that aim to make model predictions more transparent and understandable. Several techniques have been proposed to improve interpretability in radiomics and AI-based imaging models. Methods such as SHapley Additive exPlanations, feature importance analysis, and attention or saliency maps can help identify which imaging features or regions of interest contribute most strongly to a model’s predictions. These tools allow clinicians to better understand the decision-making process of AI models and to assess whether predictions align with known biological or clinical patterns. In addition, future research should emphasize biological validation of predictive imaging features, linking radiomic signatures with histopathological, genomic, or molecular characteristics of the tumor. Such integration can strengthen confidence in AI-based models and support their translation into reliable clinical decision-support systems[52-54].

Overall, recommendations for improving reproducibility and standardization in radiomics research include several practical measures. First, imaging acquisition protocols should be standardized as much as possible across institutions, including consistent slice thickness, reconstruction kernels, and contrast-enhancement timing. Second, studies should report complete technical details of imaging acquisition, preprocessing, and feature extraction pipelines to ensure transparency and reproducibility. Third, adherence to established methodological frameworks such as the IBSI and reporting guidelines such as the Radiomics Quality Score is strongly recommended. Fourth, harmonization techniques, including statistical approaches such as ComBat correction, can be used to reduce scanner-related variability in multicenter datasets. Finally, future radiomics studies should prioritize external validation using independent multicenter cohorts to confirm model generalizability before clinical implementation.

LIMITATIONS

This article has certain limitations that should be considered when interpreting the presented evidence. As a narrative review, this article provides a selective synthesis of representative studies rather than a comprehensive systematic evaluation of all available literature, which may introduce some degree of selection bias. A formal methodological quality assessment of individual studies was not performed, as the primary aim was to summarize conceptual developments in AI-driven quantitative imaging for esophageal cancer. Much of the existing evidence in this field originates from retrospective and single-center studies, which may limit the generalizability of reported findings. Variability in imaging protocols, radiomics feature extraction methods, and analytical pipelines across studies may also influence the comparability of results. In addition, the current literature may also be influenced by publication bias, as studies reporting positive predictive performance of radiomics models are more likely to be published than those reporting negative or inconclusive results. Consequently, unsuccessful model development or failed external validations may be underrepresented in the published literature. This bias may create an overly optimistic perception of the performance and clinical readiness of radiomics-based models. These considerations highlight the need for future prospective multicenter investigations and standardized methodological frameworks to strengthen the clinical applicability of AI-based quantitative imaging approaches in esophageal cancer.

CONCLUSION

The future of esophageal cancer care lies in transitioning the current state of static qualitative to static and dynamic quantitative imaging. A primary goal is the development of multi-modal fusion models that integrate radiomics and body composition with genomics and liquid biopsy data, creating a “pan-omic” profile that captures the full biological spectrum of the disease. To facilitate this, the field must prioritize the creation of large-scale, open-source, and federated databases that allow AI models to be trained on diverse global populations without compromising patient privacy. Standardizing the quantification process through automated, deep-learning-based segmentation is also essential to eliminate inter-observer bias and make real-time analysis feasible in a busy clinical workflow. Furthermore, the focus of survival modeling is expected to shift toward “delta-radiomics”, where the rate of change in tumor heterogeneity and muscle mass during the first few weeks of treatment serves as an early-warning system to switch non-responding patients to alternative therapies. Ultimately, the validation of these AI biomarkers through prospective, randomized controlled trials will be the final step in proving their clinical utility, ensuring that quantitative imaging does not just provide a more detailed map of the disease, but a more successful path to patient recovery.