Efficient Transfer Learning for Lung and Colon Histopathology: LC25000 Benchmarking with Robust Evaluation and Integrated Explainability

Automated diagnosis of cancer using digital histopathological images is a critical frontier in modern medicine. The dataset of 25,000 histological images of lung and colon cancer (LC25000), which are categorized into five distinct and uniformly distributed classes, presents the multi-class classification as a difficult challenge: Lung Adenocarcinoma, Lung Benign Tissue, Lung Squamous Cell Carcinoma, Colon Adenocarcinoma, and Colon Benign Tissue. We propose a solution with Deep Transfer Learning in EfficientNetB0, with a custom classification head and trained for 10 epochs. The model did well in an independent 10% test set of 2,500 images, providing a true Test Accuracy and a perfect F1-score for every class. A Stratified Cross-Validation of 5-fold is performed for reliability and generalizability, with an average Validation accuracy of 99%. On top of that, while 9 of the popular CNN models were evaluated in an analysis, EfficientNetB0 was the best performer. Grad-CAM imaging tests validate the model's interpretability. This validated methodology, in contrast to the results obtained by ensemble or multi-scale attention models, is a highly accurate, stable, and interpretable automated diagnostic tool that demonstrates its utility for immediate clinical evaluation.