Application of a Deep-Learning Architecture in Colon Cancer Prediction Using Histopathology Images

Colon cancer is the third most commonly diagnosed malignancy and the second leading cause of cancer-related mortality; therefore, timely and accurate diagnosis is essential. Histopathology is the gold standard, but it can be slow and variable. We present a lightweight deep learning (DL) model for binary classification of colon histopathology images that emphasises robustness and computational efficiency. Using DenseNet169 with transfer learning (TL) on 10,000 images (5,000 cancerous; 5,000 normal), we applied extensive on-the-fly augmentation and a 2-phase training strategy without stain normalization. The model achieved strong validation performance with an accuracy 0.9942, precision 0.9917, recall 0.9967, F1-score 0.9942, and AUC 0.9998, with ROC analysis indicating near-perfect separation. We utilised a three-fold cross-validation which showed consistent performance across folds, supporting generalization. Compared with stain-normalized ResNet-50 baselines, our approach remains competitive while reducing preprocessing burden, improving practicality for resource-constrained environments. These results underscore the potential of optimised TL with DenseNet169 to deliver fast, reliable decision support in pathology without complex preprocessing.