Esophageal cancer remains one of the deadliest gastrointestinal malignancies worldwide, with roughly 604,000 new cases and 544,000 deaths recorded in 2020 and five-year overall survival below 20 percent. For patients with locally advanced, resectable disease, standard care involves neoadjuvant therapy before surgery, either chemoradiotherapy established by the CROSS trial or, increasingly, chemoimmunotherapy that adds immune-checkpoint inhibitors. Yet only about 20 to 40 percent of these patients achieve a pathological complete response, the complete absence of viable tumor cells in the resected specimen. That outcome is the strongest independent predictor of long-term survival, with responders attaining five-year survival of 60 to 80 percent, while non-responders endure the cumulative toxicity of multimodality treatment without oncological benefit. A new systematic review and meta-analysis published in Frontiers in Medical Engineering now offers the first quantitative comparison of how the timing of imaging affects the ability of PET/CT radiomics to predict this pivotal outcome.
Radiomics is the high-throughput extraction of quantitative features from medical images, converting scans into hundreds of numerical descriptors that capture tumor characteristics invisible to the naked eye. When applied to 18F-fluorodeoxyglucose positron emission tomography combined with computed tomography, radiomic features include first-order statistics, texture matrices such as the gray-level co-occurrence, run-length, size-zone, and neighborhood gray-tone difference matrices, and morphological descriptors. These measurements may encode intratumoral heterogeneity that conventional semiquantitative parameters like maximum standardized uptake value, metabolic tumor volume, and total lesion glycolysis cannot capture. Because PET/CT is already integral to esophageal cancer staging, radiomics offers a non-invasive, automated route to tumor phenotyping that could transform treatment planning.
The critical insight motivating the new analysis is that radiomics can be applied at different points along the treatment trajectory, and the imaging time point is itself a biologically and methodologically important determinant of predictive performance. Baseline, pre-treatment radiomics characterizes the intrinsic biology of the untreated tumor. Restaging, post-treatment radiomics characterizes the residual metabolic phenotype but is confounded by radiation-induced inflammation that can mimic residual tumor. Delta radiomics quantifies the longitudinal change between paired scans, potentially isolating the treatment effect itself. Existing single-center studies vary widely in sample size, imaging time point, radiomic methodology, and validation strategy, and no quantitative synthesis had compared these three time points directly until now.
A team led by Jie Li and Subapriya Suppiah of Universiti Putra Malaysia, with collaborators from Beacon Hospital, Zhongshan Hospital Fudan University, Austin Health in Melbourne, and King Hussein Cancer Center in Amman, conducted the review in accordance with PRISMA-DTA guidelines and prospectively registered it with PROSPERO. The researchers searched PubMed/MEDLINE, the Cochrane Library, Scopus, and EBSCO from inception to March 2026 without language restriction. Eligible studies enrolled adults with histologically confirmed esophageal squamous cell carcinoma or esophageal adenocarcinoma scheduled for neoadjuvant chemoradiotherapy or chemoimmunotherapy followed by surgical resection, applied PET/CT radiomics at one or more time points, and used surgical histopathological examination as the reference standard for strictly defined pathological complete response.
Thirteen publications were ultimately included, generating 14 study-time-point observations, with one cohort contributing to two temporal pools. The included studies, published between 2013 and 2026, originated from the United States, the Netherlands, Taiwan, China, and Japan, with sample sizes ranging from 20 to 189 patients and reported pathological complete response rates from 20.0 to 59.2 percent. Modeling approaches ranged from univariate ROC analysis and multivariable logistic regression to support vector machines and deep-learning combined models incorporating habitat-aware radiomics and multiple-instance learning. Only two studies employed external validation, while the remainder relied on internal validation strategies such as bootstrap resampling, split-half designs, leave-one-out cross-validation, or five-fold cross-validation.
The headline quantitative result concerns the baseline pool. Eight pre-treatment studies totaling 564 patients yielded a pooled area under the receiver operating characteristic curve of 0.845, with a 95 percent confidence interval of 0.768 to 0.899 and moderate heterogeneity quantified at an I-squared of 34.6 percent. Individual study AUCs ranged from 0.74 to 0.95, with the highest estimate coming from a Japanese neural-network classifier applied to squamous cell carcinoma patients. By contrast, the four post-treatment studies totaling 387 patients produced a pooled AUC of just 0.677, with an extraordinarily wide confidence interval spanning 0.328 to 0.900 and considerable heterogeneity at I-squared of 82.8 percent. Individual post-treatment estimates ranged from 0.42 to 0.85, with the only externally validated entry achieving just 0.64, a well-recognized shrinkage of apparent performance when models are tested on new data.
The delta pool, comprising only two studies with 65 patients, was summarized descriptively rather than pooled because the minimum threshold of four studies per pool was not met. The individual delta AUCs were 0.89 and 0.79, both derived from small single-center cohorts using texture and histogram-distance features. The authors note that an illustrative random-effects interval for the delta pool would span 0.04 to 1.00, rendering any quantitative pooled estimate uninformative. They characterize the delta findings as hypothesis-generating only, though conceptually attractive because delta radiomics isolates treatment-induced change in tumor phenotype and may be less sensitive to inter-scanner differences in absolute feature values.
An exploratory subgroup analysis restricted to baseline estimates compared radiomics-only models with models combining radiomics and clinical variables. The radiomics-only subgroup of six studies yielded a pooled AUC of 0.831, while the radiomics-plus-clinical subgroup of four studies achieved 0.814, a negligible between-subgroup difference of 0.017 that was not statistically significant. However, within-cohort paired analyses told a different story: in individual studies, adding clinical variables increased the AUC from 0.71 to 0.852 and from 0.772 to 0.824. The authors caution that with so few studies, these exploratory results should be interpreted with care, and publication bias could not be reliably assessed because fewer than 10 studies were available in each pool, precluding funnel-plot analysis in accordance with the Cochrane Handbook.
Methodological quality assessment revealed both reassurance and concern. Under the QUADAS-2 tool, eight study entries achieved an overall low risk of bias and five were rated unclear, with no entry rated high risk. The most frequent source of uncertainty was the index-test domain, reflecting infrequent reporting of blinding of radiomic model development and segmentation to the pathological outcome, together with thresholds derived from the same dataset used for model training. The Radiomics Quality Score, which assesses radiomics-specific dimensions across 16 components with a maximum of 36 points, was moderate overall with a median of 44.4 percent and individual scores from 19 to 67 percent. Recurrent deficiencies included phantom and inter-observer reproducibility testing, compliance with the Image Biomarker Standardization Initiative, prospective registration, and open-science practices, shortcomings that could prevent reliable deployment in multicenter practice.
The findings carry direct clinical implications for organ-preservation strategies. The preSANO and SANO trials demonstrated that selective non-operative management can be oncologically safe in patients with clinical complete response after neoadjuvant chemoradiotherapy, creating an urgent need for reliable pre-treatment biomarkers to triage candidates. A baseline radiomic model with consistent accuracy near 0.85 could complement endoscopic and clinical assessment to stratify patients before therapy, intensifying treatment in predicted non-responders and prioritizing predicted responders for closer surveillance, potentially in combination with circulating tumor DNA monitoring. The emergence of chemoimmunotherapy adds another dimension: three baseline-pool studies in the immune-checkpoint-inhibitor setting achieved high accuracy with AUCs of 0.852, 0.87, and 0.824, and the highest pathological complete response rates in the dataset, consistent with the additive antitumor effect of PD-1 blockade. The authors suggest that features of the tumor immune microenvironment driving checkpoint response, such as CD8-positive T-cell infiltration and PD-L1 expression, may be partially encoded in pre-treatment FDG PET radiomic features through their association with metabolic activity and intratumoral heterogeneity.
Important caveats temper these encouraging signals. All included studies were retrospective and predominantly single-center, with only two employing external validation. The baseline and post-treatment estimates were derived from different study sets and should not be interpreted as a direct head-to-head comparison. Acquisition protocols differed across studies in injected dose, uptake time, and acquisition duration, potentially affecting feature comparability, and each pooled AUC summarizes methodologically diverse models rather than a single standardized signature. The authors conclude that multicenter prospective studies employing standardized, IBSI-compliant radiomic workflows, which compare baseline, delta, and post-treatment imaging directly within the same cohorts, represent the critical next step toward clinical translation, alongside dedicated model development for chemoimmunotherapy regimens, integration with liquid-biopsy and tissue biomarkers, and health-economic evaluation of radiomic-guided treatment stratification.
Subject of Research: Diagnostic accuracy of 18F-FDG PET/CT radiomics across baseline, restaging, and delta imaging timepoints for predicting pathological complete response to neoadjuvant therapy in locally advanced esophageal cancer
Article Title: 18F-FDG PET/CT radiomics across treatment timepoints for predicting response to neoadjuvant therapy in locally advanced esophageal cancer: a systematic review and meta-analysis
Article References: Li, J., Suppiah, S., Kasilingam, L., Hui, S. T., Shi, H., Lee, S. T., Sharin, M. F. B. M., & Al-Ibraheem, A. (2026). 18F-FDG PET/CT radiomics across treatment timepoints for predicting response to neoadjuvant therapy in locally advanced esophageal cancer: a systematic review and meta-analysis. Frontiers in Medical Engineering, 4, Article 1979078. https://doi.org/10.3389/fmede.2026.1979078
Image Credits: AI Generated
DOI: 10.3389/fmede.2026.1979078
Keywords: 18F-FDG PET/CT, radiomics, esophageal cancer, pathological complete response, neoadjuvant therapy, meta-analysis, systematic review, delta radiomics, immune-checkpoint inhibitors, machine learning, treatment response prediction, medical imaging
News Source: Nathaniel Bowman. (October 16, 2026). PET/CT Radiomics Before Treatment Best Predicts Esophageal Cancer Therapy Response, Meta-Analysis Finds. Scienmag.
