Detecting aberrant p53 immunohistochemical expression patterns in patients with Barrett’s esophagus using artificial intelligence

Abstract

Purpose: Immunohistochemistry (IHC) for the tumor suppressor protein p53 is an adjunct biomarker for Barrett’s esophagus (BE)-related dysplasia classification and risk stratification. Four phenotypic staining patterns are distinguished: wild-type (WT), representing normal staining, and three aberrant patterns: overexpression (OE), null mutation (NM), and double clone (DC). OE is readily recognized; NM and DC are harder to identify and often overlooked. This diagnostic challenge leads to variation in patient treatment because DC is associated with higher progression rates. An AI tool could improve diagnostic accuracy and patient risk stratification. Approach: We compared two AI approaches discriminating p53 patterns: a full biopsy (FB) approach and a clustering-constrained-attention multiple-instance learning (CLAM) approach. Because DC is rare, we adopted a double-binary strategy (CLAM-DB, FB-DB) that predicts OE and NM separately, and we introduced synthetic DC (CLAM + SynDC) by combining OE and NM patterns to boost DC performance. DC cases were excluded from training and used only for testing. Results: For aberrant versus non-aberrant expression, all models performed similarly (AUC 0.94 to 0.96). For four-class p53 IHC (WT, OE, NM, DC), CLAM-based models consistently outperformed FB models in overall accuracy. CLAM + SynDC and CLAM-DB outperformed baseline CLAM (0.84, 0.84 versus 0.71), whereas combining DB and SynDC did not yield further gains. Conclusion: Double-binary training enabled robust recognition of DC despite its absence during training, underscoring the value of disentangling OE and NM signals. Overall, AI can provide a standardized and clinically relevant aid for interpreting p53 IHC in BE and supporting patient risk stratification. Code is available at GitHub https://github.com/qurAI-amsterdam/BE_p53_classification.