Multiclass characterization of colorectal polyps under class imbalance using a calibrated cascade model

Colorectal cancer (CRC) is a significant global health issue, responsible for numerous annual cancer-related deaths. Colorectal polyps (CRPs), as precursors to CRC, necessitate early detection and precise characterization to enhance patient outcomes and decrease mortality rates. While much of the current research focuses on benign-malignant classification, the advancement of multi-class characterization for polyp types through computer-assisted analysis is gaining critical importance in enhancing diagnostic accuracy and supporting clinical decision-making. However, a significant challenge lies in addressing class imbalances within the data, which can bias model performance towards more prevalent classes and hinder the detection of less common but potentially more dangerous polyps. To address this challenge, this study introduces a sequential binary decision-making approach for characterizing CRP pathologies, distinguishing between Adenocarcinoma, Adenoma, and Hyperplasia. This method aims to leverage the structured decision-making advantages of decision trees within a neural network-based approach by decomposing the complex task of multi-class characterization into a sequential series of binary decisions. Each binary classifier focuses on distinguishing a specific class (one-vs.-all), enabling a more interpretable decision-making process. When combining this approach with calibration, the resulting performance demonstrates that the proposed calibrated cascade model achieves notable improvements over conventional multi-class CNN models and ensemble approaches, with a 2.8% improvement in F₁ score compared to the state-of-the-art method. By addressing class imbalance and incorporating confidence calibration, this approach offers a reliable and interpretable solution for multi-class