High-Accuracy Multi-Class Lung and Colon Histopathology Classification Utilizing Deep Transfer Learning with EfficientNet-B7

Abstract

Automated histopathological analysis plays an increasingly vital role in modern oncological diagnostics, offering objective decision support alongside clinical pathology workflows. In this study, we present a deep transfer learning framework leveraging the scaled EfficientNet-B7 architecture to perform unified five-class classification across lung and colon tissues using the LC25000 histopathological image dataset. The model is trained on 25,000 hematoxylin and eosin (H&E) stained tissue images covering five categories: colon adenocarcinoma, benign colonic tissue, lung adenocarcinoma, benign pulmonary tissue, and lung squamous cell carcinoma. By freezing the ImageNet-pretrained compound-scaled feature extractor and appending a dense multi-layer classification head, the framework converges rapidly using Adam optimization and categorical cross-entropy loss. Evaluated across an independent validation cohort of 5,000 images, the proposed model achieves a peak validation accuracy of 99.10% alongside an overall macro-averaged precision, recall, and F1-score of 0.991. The normalized confusion matrix exhibits 100% sensitivity for benign lung and colon tissues, confirming strong diagnostic specificity with minimal misclassification confined primarily to overlapping non-small cell lung carcinoma subtypes. These results validate the efficacy of compound-scaled deep feature representations as an accurate, robust solution for automated multi-organ cancer screening.