Evaluation of pancreatic cancer fibrosis based on CT images and its association with chemotherapy efficacy: A multicenter study.

Abstract

e16444 Background: Tumor fibrosis significantly impacts chemotherapy efficacy in pancreatic ductal adenocarcinoma (PDAC), but its prognostic role remains unclear. This study aims to develop an artificial intelligence (AI)-based model for non-invasive fibrosis prediction and explore its relationship with chemotherapy efficacy. Methods: A three-phase study was conducted, including fibrosis quantification via whole slide images (WSIs), non-invasive fibrosis prediction using preoperative CT images, and evaluation of its association with chemotherapy outcomes. The surgical cohort included 123 WSIs from The Cancer Genome Atlas (TCGA), 150 WSIs with corresponding CT images from Sun Yat-sen University Cancer Center (SYSUCC), and 87 WSIs with CT images from Xiangya Hospital of Central South University (XHCSU). Additionally, we collected baseline CT images and survival data from 296 chemotherapy-treated PDAC patients at SYSUCC, grouped by AG, FOLFIRINOX, or SOXIRI regimens. A deep learning-based tissue recognition model was trained on WSIs to quantify stromal tissue proportions as a surrogate for fibrosis extent. The median stromal proportion from the SYSUCC cohort was used to define high- and low-fibrosis groups, which were subsequently used to train the fibrosis prediction model based on CT images. CT data from SYSUCC were used for training, and XHCSU served as an external validation set. Results: The tissue recognition model achieved eight-class accuracies of 99%, 97%, and 95% across cohorts. High fibrosis was significantly associated with improved overall survival in all cohorts (log-rank test P values = 0.044 for SYSUCC, 0.032 for TCGA, and 0.029 for XHCSU). The fibrosis prediction model achieved AUCs of 0.736 in training and 0.706 in external validation. Using the predicted fibrosis groups, we found that high-fibrosis patients receiving the AG regimen showed significantly improved outcomes (log-rank test P values = 0.001 for overall survival and 0.032 for progression-free survival), while no significant survival differences were observed between fibrosis groups for FOLFIRINOX or SOXIRI. Conclusions: This study presents a robust AI-based method for non-invasive fibrosis assessment. Our findings highlight that fibrosis extent is a critical determinant of chemotherapy efficacy, with high-fibrosis patients benefiting most from AG. This approach has great potential to aid clinicians in tailoring chemotherapy regimens based on non-invasive fibrosis prediction, ultimately improving outcomes for PDAC patients.