P1.123. Deep Learning and Machine Learning Radiomics-Based Models Predicting Esophagogastric Cancer Recurrence: A Systematic Review

Abstract

Abstract Topic Esophageal Cancer: Oncology/Radiation Therapy Background Esophageal and gastric cancers are associated with recurrence rates approaching 50% despite curative-intent treatment. Non-invasive imaging modalities including computed tomography (CT), positron emission tomography (PET/CT), and magnetic resonance imaging (MRI) are routinely used for staging and surveillance. Radiomics combined with machine learning (ML) and deep learning (DL) enables extraction of high-dimensional imaging features that may predict recurrence risk. However, clinical translation is limited due to the heterogeneity in methodology, validation strategies, and reporting standards limits. A comprehensive evaluation of performance, quality, and reproducibility of these models is currently lacking. Methods A systematic review was conducted following PRISMA guidelines. Searches were performed in PubMed, Embase, Medline, Cochrane Library, and additional databases using predefined terms related to esophageal adenocarcinoma, esophagogastric junction adenocarcinoma, gastric cancer, radiomics, machine learning, deep learning, CT, PET, and MRI. Primary studies evaluating radiomics-based ML or DL models predicting local, distant, or metastatic recurrence were included. Studies limited to conventional metrics (e.g., tumor volume, SUV, MTV), animal or phantom studies, reviews, and conference abstracts were excluded. Two reviewers independently screened titles, abstracts, and full texts using Covidence. Data extraction was piloted and includes study design, imaging protocol, feature extraction methods, model development, validation strategy, and performance metrics. Methodological quality and reporting will be assessed using Radiomics Quality Score (RQS) and Quality Assessment of Diagnostic Accuracy Studies (QUADAS). Results Database screening has been completed and eligible studies identified. Data extraction has been piloted to ensure consistency and feasibility of planned analyses. Most studies appear to be retrospective and single-center, with variable use of external validation. Planned quantitative synthesis will summarize model performance metrics including sensitivity, specificity, and area under the receiver operating characteristic curve, results will be available by the end of June, given that the abstract is accepted. Where appropriate, pooled estimates with 95% confidence intervals will be calculated. Radiomics Quality Score adherence and QUADAS risk of bias assessments will be systematically summarized to evaluate methodological robustness and translational potential. Conclusion Radiomics-based ML and DL models show promise for predicting esophagogastric cancer recurrence; however, methodological variability and reporting limitations may hinder generalizability. This systematic review will provide a structured appraisal of model performance, quality, and risk of bias, identifying key gaps and priorities for future research. The findings aim to support development of reproducible, externally validated models suitable for clinical implementation in recurrence risk stratification.