Treatment-specific CT Radiomics Models To Predict Response To Neoadjuvant Therapy And Explore Individualized Treatment Selection In Gastric Cancer.

RATIONALE AND

OBJECTIVES

To develop and compare general and treatment-specific radiomics models based on pretreatment computed tomography (CT) for predicting pathological response to neoadjuvant therapy (NAT) in gastric cancer (GC), and to explore a dual-score framework for individualized treatment selection.

MATERIALS AND

METHODS

This retrospective study included 405 patients with GC who underwent neoadjuvant chemotherapy (NAC) or neoadjuvant immunochemotherapy (NAIC) followed by radical gastrectomy, comprising 337 in the development cohort and 68 in a temporal test cohort. The development cohort was randomly divided into training (n = 235) and validation (n = 102) sets. Radiomics features were extracted from portal venous-phase CT images. Four machine learning classifiers were used to construct general and treatment-specific models. Treatment-specific models were cross-applied to generate paired NAC and NAIC response probabilities.

RESULTS

In validation, the general model achieved an AUC of 0.679 (NAC, 0.732; NAIC, 0.659), whereas the NAC-specific and NAIC-specific models achieved AUCs of 0.770 and 0.753. Repeated random-split analyses more frequently favored treatment-specific models. In the temporal test cohort, the NAIC-specific model outperformed the general model (AUC, 0.707 vs 0.626), whereas the NAC-specific model showed no advantage (AUC, 0.563 vs 0.625). In the dual-score framework, patients who received model-recommended treatment showed higher pathological response rates (NAC-recommended: 46.4% vs 23.3%, p = 0.043; NAIC-recommended: 40.5% vs 19.4%, p < 0.0001).

CONCLUSION

Treatment-specific radiomics models showed better discrimination than the general model for predicting pathological response to NAT in gastric cancer. The dual-score framework may provide an exploratory approach for individualized treatment selection.