TRACE-CLDN: An Integrated Computational Pathology Framework for Tumor-Restricted Quantification of CLDN18.2
Yu-Meng Pan1, Hao-Tian Zhang1, Meng-Ling Liu1
1School of Health Science and Engineering, University of Shanghai for Science and Technology, Shanghai, 200093, China.
Abstract:
Claudin-18.2 (CLDN18.2) is a membrane-associated biomarker with emerging translational relevance in biliary tract cancers. In intrahepatic cholangiocarcinoma (ICC), whole-slide image (WSI)-level assessment is complicated by abundant stroma, dispersed tumor glands, non-neoplastic epithelial structures, and heterogeneous membranous staining. We developed TRACE-CLDN (Tumor-Restricted Analysis for Computational Estimation of CLDN18.2), an integrated computational pathology framework for tumor-restricted quantification of CLDN18.2 expression burden from hematoxylin and eosin (H&E) WSIs. This retrospective study included 135 ICC cases with paired H&E and CLDN18.2 immunohistochemistry (IHC) WSIs. After cross-modal registration, IHC-derived annotations were mapped to H&E images to generate patch-level training data. TRACE-CLDN integrates two independently trained H&E-based modules for tumor region identification and CLDN18.2 expression label prediction, followed by WSI-level tumor-restricted aggregation. Agreement with pathologist consensus was assessed in a 45-case pathologist-validation subset, and exploratory survival analysis was performed in the full cohort. ROC-AUCs were 0.9968 ± 0.0009 for tumor identification and 0.9633 ± 0.0174 for expression prediction. The TRACE-CLDN-derived Ratio showed strong agreement with pathologist assessment (Spearman ρ = 0.876; weighted κ = 0.767). In exploratory survival analysis, the log-rank P value was 0.010 and the concordance index was 0.668. TRACE-CLDN enables tumor-restricted, continuous assessment of CLDN18.2 expression burden from H&E WSIs, with strong agreement with pathologist assessment and an exploratory association with survival in ICC.
