Construction of a CT Radiomics Model for Predicting Her2 Expression in Bladder Cancer Based on Random Forest Algorithm

Abstract

Objective This study aimed to develop and evaluate a predictive model for Human Epidermal Growth Factor Receptor 2 (HER2) expression levels in bladder cancer patients using clinical data and computed tomography (CT) radiomic features across various imaging phases. Methods The investigation involved: (1) compiling clinical data from bladder cancer patients; (2) performing HER2 immunohistochemistry (IHC) assessments post-surgery using the Hercep Test scoring system; (3) delineating tumor regions on CT images to extract radiomic features; (4) utilizing T-tests and Least Absolute Shrinkage and Selection Operator (LASSO) regression to identify the most predictive radiomic features of HER2 status. Decision trees and random forest algorithms were then employed to construct radiomic models. Each model's predictive accuracy, sensitivity, specificity, and area under the curve (AUC) were evaluated through cross-validation, identifying the model with the highest AUC as the optimal radiomic predictor. Results The study included 84 bladder cancer patients, with 53 classified as HER2-negative and 31 as HER2-positive via IHC. Radiomic features that correlated with HER2 status were identified, with three, eight, and two features selected from non-contrast, arterial, and venous phase CTs, respectively. Models based solely on arterial phase features exhibited modest predictive capacity (AUC = 0.44), which improved slightly with the inclusion of clinical data (AUC = 0.48). However, a model integrating features from all three CT phases (totaling 13 features) significantly enhanced performance, achieving an AUC of 0.78, which further improved to 0.83 when combined with clinical variables. Conclusion CT-based radiomics is a viable method for predicting HER2 expression in bladder cancer. The comprehensive model, incorporating features from non-contrast, arterial, and venous CT phases alongside clinical data, demonstrated superior predictive efficacy.