EP245 - Predicting the Effectiveness of Neoadjuvant Therapy in Rectal Cancer Patients: A Model Based on CT Radiomics and Carcinoembryonic Antigen

Purpose/Background: Neoadjuvant chemoradiotherapy (NCRT) is standard for locally advanced rectal cancer (LARC), but treatment response varies significantly. While MRI-based predictive models are common, CT is more widely available and convenient. This study aims to develop and validate a multimodal model integrating CT radiomics features and carcinoembryonic antigen (CEA) levels to predict the efficacy of NCRT in rectal cancer patients.

Methods/Interventions: Data were retrospectively obtained from 183 patients with locally advanced rectal cancer who underwent neoadjuvant chemoradiotherapy (NCRT) and subsequent radical surgery. Patients were stratified into good (TRG 0-1) and poor (TRG 2-3) response groups based on post-operative pathology. Radiomics features were extracted from pre-NCRT pelvic enhanced CT images using 3D-Slicer for tumor segmentation and GE's A.K. software for feature calculation. CEA levels were collected and standardized. After feature selection via LASSO regression, predictive models were constructed using Logistic Regression, Support Vector Machine, Random Forest, and XGBoost algorithms. These models were developed and compared in three categories: a clinical model, a pure radiomics model, and an integrated model combining both data types. Model performance was evaluated on a held-out testing set (30% of patients) using the Area Under the Curve (AUC), accuracy, sensitivity, and specificity.

Results/Outcomes: In conclusion, our study successfully developed and validated a multimodal model that effectively predicts the response to NCRT in LARC patients by integrating CT-based radiomics features with the serum tumor marker CEA. The integrated model, particularly the XGBoost algorithm, achieved the best predictive performance (AUC of 0.98 in the testing set), significantly surpassing both the pure clinical and pure radiomics models. This underscores the complementary value of combining imaging phenotypes with key clinical biomarkers for a more robust assessment.

Our findings demonstrate the clinical viability of CT-based radiomics, offering a widely available and practical alternative to MRI-centric approaches. This model can serve as a valuable non-invasive tool to identify patients likely to benefit from NCRT before treatment initiation, thereby aiding in personalized therapeutic decision-making. While promising, further multi-center, prospective studies with larger cohorts are warranted to solidify its generalizability and prepare for clinical integration.

Conclusion/Discussion: We successfully developed a predictive model that combines CT radiomics and clinical factor CEA. The integrated model, particularly using the XGBoost algorithm, shows excellent performance in predicting NCRT response for rectal cancer. This tool holds promise for assisting clinicians in personalizing treatment strategies. Further prospective validation is warranted to assess its real-world clinical utility.

Performance Analysis of Radiomics Fusion Models Using ROC and Calibration Curves