Deep Fusion: A High‐Performance AI Framework for Colorectal Cancer Grading

ABSTRACT Automated, precise histopathological grading of colorectal cancer (CRC) is vital for prognosis and treatment but is challenged by inter‐observer variability and time demands. This study introduces and evaluates a novel deep learning framework for robust four‐class grading of colorectal adenocarcinoma (Normal, Well, Moderately, and Poorly Differentiated), directly addressing limitations of prior studies focused on survival prediction or limited feature extraction. Our approach integrates Reinhard color normalization for stain variability, data augmentation for robustness, and synergistic feature extraction from fine‐tuned MobileNetV2 and InceptionV3 Convolutional Neural Networks (CNNs). Computationally efficient classification was achieved using Principal Component Analysis (PCA) to reduce features to 100 components, followed by an optimized Support Vector Machine (SVM) with multiple kernel evaluations. The framework, with a polynomial kernel SVM, demonstrated superior performance, achieving 96.41% macro‐averaged accuracy (95% CI, 95.8%–97.0%), with corresponding 0.9641 precision, recall, F 1‐score, and a 0.9915 macro area under the curve (AUC) in fivefold cross‐validation. Polynomial and Radial Basis Function (RBF) kernels significantly outperformed others. The framework's primary contribution is a validated, synergistic feature fusion pipeline that leverages the complementary strengths of diverse CNN architectures and an optimized SVM classifier, representing a notable advancement in automated grading on clinically sourced data. This provides a comprehensive framework for multi‐class CRC grading, showing considerable potential to enhance diagnostic accuracy, consistency, and efficiency.