This study introduces a novel transfer learning-based contrastive language-image pretraining adapter (CLIP-adapter) model for predicting the tumor-stroma ratio (TSR) in pancreatic ductal adenocarcinoma (PDAC) using preoperative dual-phase CT images. The primary aim is to develop an efficient and accessible tool for risk stratification and personalized treatment planning.

The proposed novel Contrastive Language-Image Pretraining-Adapter (CLIP-adapter) model, leveraging transfer learning, framing CLIP and a self-attention mechanism for predicting TSR in PDAC, in order to exhibit high performance in distinguishing low and high TSR PDAC in the test cohort. We speculated the CLIP-adapter model outperformed single-phase models, specifically CLIP models based on arterial or venous phase images alone. The addition of a feature fusion module could enhance the model's differentiation capacity, emphasizing its superiority over single-phase models. Besides, the model we designed utilized both image and text information during network training, instead of focusing on images only. This underscores the importance of comprehensive assessment in PDAC imaging evaluation, with the potential to contribute to risk stratification and personalized treatment planning.

Inclusion Criteria:

1. patients with pathologically proven PDAC by surgical resection
2. patients who underwent CT scan within a month before surgery
3. observable pancreatic lesions on available images.

Exclusion Criteria:

1. any anti-cancer therapy before CT scanning
2. conspicuous interference or significant motion distortions found on images
3. partial clinical data
4. patients with liver metastases or peritoneal carcinomatosis prior to surgical intervention.

A total of 207 patients were chosen from three independent4 hospitals: the First Affiliated Hospital of Chongqing Medical University (FAHCQMU), Daping Hospital of Army Medical University (DPHAMU), and the Third Affiliated Hospital of Chongqing Medical University (TAHCQMU).