Multicenter prospective validation of PanClaudinAI: An AI for predicting Claudin 18.2 from CT in pancreatic cancer.

Abstract

751 Background: Claudin 18.2 (CLDN18.2) is a promising therapeutic target in pancreatic cancer. Current immunohistochemistry (IHC)-based assessment is invasive and limited by tumor heterogeneity. We developed a deep learning model using contrast-enhanced CT (CECT) to predict CLDN18.2 expression and validated it in a multicenter prospective cohort via a dedicated software application (PanClaudinAI). Methods: We retrospectively collected CECT images and matched CLDN18.2 IHC data from 800 patients across three centers (Center A: n = 400; Center B: n = 230; Center C: n = 170). CLDN18.2 positivity was defined as ≥75% moderate-to-strong staining. A Vision Transformer (ViT) model was trained using arterial-phase CT images. The model was integrated into a user-friendly software application (PanClaudinAI) for real-time inference. We subsequently conducted a prospective multicenter validation (Center D: n = 100; Center E: n = 120; Center F: n = 140) to evaluate its clinical utility. Results: The prevalence of CLDN18.2 positivity was 45.7% in the retrospective cohort and 50.2% in the prospective cohort. The AI model achieved an AUROC of 0.81 (95% CI: 0.76–0.86) with a sensitivity of 78.2% and specificity of 75.4% in center A; an AUROC of 0.84 (95% CI: 0.81–0.89) with a sensitivity of 81.5% and specificity of 79.3% in center B; and an AUROC of 0.86 (95% CI: 0.83–0.89) with a sensitivity of 82.7% and specificity of 80.1% in center C within the retrospective training set. In the prospective multicenter validation, the model maintained an AUROC of 0.79 (95% CI: 0.73–0.84) with a sensitivity of 76.5% and specificity of 74.2% in center D; an AUROC of 0.78 (95% CI: 0.74–0.82) with a sensitivity of 75.8% and specificity of 73.6% in center E; and an AUROC of 0.85 (95% CI: 0.82–0.91) with a sensitivity of 81.2% and specificity of 79.0% in center F. Moreover, the PanClaudinAI software demonstrated high usability and integration into clinical workflow, with an average processing time of < 2 minutes per case. Conclusions: We developed and prospectively validated a robust AI-based software (PanClaudinAI) that non-invasively predicts CLDN18.2 expression from routine CECT images across multiple centers. This tool facilitates rapid, reproducible, and accessible biomarker identification, potentially guiding patient selection for CLDN18.2-targeted therapies and reducing the need for invasive biopsies.