CT imaging-based radiomics and deep learning models for predicting chemotherapy response in advanced pancreatic cancer
Zhu Y, H J Zhou, Peng An, Yingfan Mao, Ziwei Nie, Yi-Xiang Wang, Zi Wang, Wenjing Cui
To investigate the value of radiomics and deep learning features derived from pre-treatment CT imaging in predicting the efficacy of chemotherapy in patients with advanced pancreatic cancer. The retrospective study included 207 patients with advanced pancreatic cancer from two medical centers, divided into a training cohort ( n = 139) and an external validation cohort ( n = 68). Radiomics features were extracted from portal-venous-phase CT images and selected via Least Absolute Shrinkage and Selection Operator (LASSO) regression. A variational autoencoder (VAE) extracted deep learning features from three-dimensional tumor volumes. The clinical, radiomics, and combined models were constructed and compared using the DeLong test. Twenty-three radiomics features were selected for model construction. In the training cohort, the radiomics model (AUC = 0.953) and the combined model (AUC = 0.965) outperformed the clinical model (AUC = 0.765, P < 0.001). In the external validation cohort, the radiomics model achieved an AUC of 0.982, with a precision of 0.85, a recall of 0.83, and an F1-score of 0.84 for identifying responders. Multivariate analysis identified cT and cN stages as independent predictors of treatment response. VAE-based deep features maintained prediction accuracy of 0.75–0.82 across different classifier types. CT-based radiomics models can predict chemotherapy response in advanced pancreatic cancer, outperforming clinical models. VAE-derived deep learning features also show promising predictive performance.