Lung and Colon Cancer Detection using Hybrid Deep Learning Models

Early and accurate detection of lung and colon cancer remains one of the most critical challenges in modern pathology, particularly when relying on manual interpretation of histopathological images. This paper proposes a fully automated diagnostic framework based on deep learning for binary cancer classification using the LC25000 dataset. All histopathological images were preprocessed with resizing to 224×224 pixels, random horizontal flipping, and random rotation augmentation. The task was formulated as binary classification: cancerous tissue versus normal tissue. Six deep learning architectures were benchmarked under identical training conditions: EfficientNet-B0, Vision Transformer (ViT-B16), DenseNet121, ConvNeXt-Tiny, MobileNetV3-Large, and ResNet34. EfficientNet-B0 achieved the highest individual accuracy of 99.76% with F1 score of 0.9865, while ViT-B16 reached 99.47% accuracy with F1 score of 0.9698. Based on these results, both models were selected for hybridization. The proposed Hybrid EfficientNet–ViT architecture, which concatenates features from both backbones into a single linear classifier, was trained for 15 epochs using Adam optimizer with mixed-precision on a T4 GPU. The hybrid model achieved 99.95% accuracy, 100% precision, 99.47% recall, 99.73% F1 score, 100% specificity, and ROC-AUC of 0.9999986, outperforming all individual baselines. Grad-CAM explainability is integrated to provide visual interpretability for clinical use.