Evaluating CNN Architectures and Transfer Learning for Histopathological Classification of Lung and Colon Cancer

Abstract

Lung and colon cancers are among the most prevalent and fatal diseases worldwide, characterized by high morbidity and mortality rates. In recent years, several state-of-the-art convolutional neural network (CNN) architectures, including VGG19, ResNet50, InceptionV3, MobileNetV2, and EfficientNet-B4, have demonstrated remarkable performance in histopathological image analysis and cancer classification. Expanding on these developments, this study aims to classify lung and colon cancers based on microscopic pathology images. To this end, six different datasets (A–F) were derived from the publicly available LC25000 dataset. Various data augmentation techniques were applied, and image dimensions were standardized to ensure consistency. Multiple CNN-based models were developed and evaluated to distinguish between three classes of colon and lung cancer. Additionally, the effectiveness of transfer learning approaches was examined due to the limited number of available samples. Experimental results indicated that the EfficientNet-B4 model achieved the highest classification accuracy of 95.08%. However, despite the high accuracy, relatively lower recall, F1-score and precision values in certain cases suggest limitations in the model’s ability to consistently identify positive cases and capture all relevant examples within the dataset. These findings provide valuable insights into the strengths and weaknesses of modern CNN architectures for histopathological cancer classification and emphasize the importance of balancing accuracy with sensitivity and precision in clinical diagnostic applications.