Abstract
Colon cancer continues to be a major contributor to cancer-related deaths worldwide, highlighting the critical need for reliable and early detection methods. In response, this research introduces an advanced deep learning framework for the automated identification of colon cancer through histopathological image analysis. The framework integrates Convolutional Neural Networks (CNNs) with Bayesian optimization to efficiently fine-tune hyperparameters, enhancing classification accuracy while minimizing overfitting. The model was trained and tested using a publicly accessible dataset that merges data from Kaggle and the Kaggle Cancer Data Portal (KCDP), covering nine distinct tissue types: Normal, Tumor, Stroma, Lympho, Complex, Debris, Mucosa, Adipose, and Background. The optimized CNN demonstrated strong performance, achieving an accuracy of 96.84%, a precision of 97.02%, a recall of 96.50%, and an F1-score of 96.71%. Additionally, the model attained an AUC (Area Under Curve) of 0.97, indicating high discriminative capability. Compared to baseline CNN and ResNet architectures, the proposed method demonstrated superior robustness and generalization due to effective data augmentation and stain normalization techniques. These findings suggest that the model offers substantial promise as a computer-aided diagnosis (CAD) tool to assist pathologists in clinical decision-making, and can be extended to other cancer types through transfer learning and model adaptation. While the framework demonstrates strong within-dataset performance, external validation on independent, multi-institutional cohorts is required before clinical deployment.