Optimized Deep Ensemble Framework for Colorectal Polyp Detection and Clinical Deployment Design

Abstract

Early and accurate detection of colorectal polyps is critical for reducing colorectal cancer risk and improving patient outcomes. This paper introduces an ensemble deep transfer learning framework with Bayesian hyperparameter optimization for robust colorectal polyp classification. The method combines three state-of-the-art backbones—ResNet50, EfficientNetB0, and InceptionV3—whose outputs are fused via probability averaging to improve reliability. Stratified 10-fold cross-validation provides unbiased performance estimates, while Bayesian optimization fine-tunes model parameters for high accuracy and efficiency. Experiments on two benchmark datasets demonstrate excellent results, achieving 99.56% accuracy on CP-CHILD-A and 99.40% on CP-CHILD-B. To illustrate clinical usability, we also designed user interface prototypes as a computer-aided diagnostic (CAD) system, showing how the framework could be integrated into real-world screening workflows. These results highlight the potential of the proposed approach for real-time, clinically deployable colorectal polyp detection.