Histopathological Identification of Lung and Colon Combine Cancer using Deep Learning Approach

Abstract

Accurate histopathological differentiation of lung and colon malignancies is essential for timely clinical intervention, yet visual overlapping among tissue patterns often hinders manual diagnosis. This study presents a deep-learning framework that leverages a fine-tuned VGG-16 convolutional neural network to classify combined lung and colon cancer images from the publicly available LC25000 dataset. The dataset comprises 25,000 hematoxylin-and-eosin–stained micrographs evenly distributed among benign and malignant categories for both organs. ImageNet pre-trained weights were transferred, and the terminal convolutional blocks were unfrozen for domain-specific training while extensive rotation, flipping, and color-jitter augmentation mitigated class imbalance and overfitting. Following Adam optimization with an initial learning rate of 1 × 10-4, the model converged in 30 epochs, achieving 95 % overall accuracy, 0.94 F1-score, and 0.96 area under the receiver-operating characteristic curve on a held-out test set. Saliency and Grad-CAM visualizations indicated that discriminative regions corresponded to nuclear pleomorphism and glandular architecture, thereby providing interpretability consistent with pathological criteria. The experimental outcomes demonstrate that transfer learning with VGG-16 yields a fast, resource-efficient, and highly reliable computer-aided diagnostic tool for large-scale screening of lung and colon histopathology slides. Future enhancements may incorporate transformer-based attention and ensemble strategies to further refine performance across heterogeneous staining protocols.