Combined dual-phase CT radiomics and deep transfer learning model for oxaliplatin response in unresectable colorectal...

Accurate prediction of response to first-line oxaliplatin-based chemotherapy in unresectable colorectal liver metastases (CRLM) is critical for optimizing treatment strategies. This study aimed to develop and validate a dual-phase CT-based Combined model integrating radiomics, deep transfer learning (DTL), and clinical variables to enable individualized response prediction.

In this multicenter retrospective study, 360 patients with unresectable CRLM were enrolled from three medical centers. Patients from one center were randomly divided into a training cohort (n = 199) and an internal validation cohort (n = 86), while an independent external test cohort (n = 75) was composed of patients from two separate institutions. Radiomics and DTL features were extracted from arterial and portal venous phase CT images. A Combined model was developed by integrating radiomics, DTL outputs, and clinical variables. Model performance was evaluated through AUC, accuracy, specificity, sensitivity, and clinical utility.

The Combined model achieved AUCs of 0.928 (95% CI: 0.894-0.963), 0.899 (95% CI: 0.834-0.964), and 0.858 (95% CI: 0.776-0.941) in the training, internal validation, and external test cohorts, respectively. It consistently outperformed the radiomics, DTL, and Clinical models in most cohorts, with favorable calibration and greater clinical benefit across decision thresholds.

We developed and validated a dual-phase CT Combined model that enables accurate early prediction of chemotherapy response in unresectable CRLM. By facilitating early identification of patients with non-objective response, this approach provides a non-invasive and clinically feasible tool to support individualized treatment planning, guide timely treatment modification, and ultimately improve patient outcomes.