Lung cancer classification using vision transformer: a CRISP-DM approach with histopathological imaging

Abstract

Lung cancer classification based on histopathological imaging plays a pivotal role in achieving early detection, accurate diagnosis, and effective treatment planning. Conventional diagnostic methods, including manual examination of histopathological slides and radiological imaging, are often subjective and time-consuming. Their limited ability to capture complex morphological patterns further constrains diagnostic accuracy. In response to these limitations, the present study employs the Cross-Industry Standard Process for Data Mining (CRISP-DM) framework to systematically evaluate the application of the Vision Transformer (ViT) in lung cancer classification. The LC25000 dataset, comprising three histopathological categories: adenocarcinoma, squamous cell carcinoma, and benign lung tissue. It was utilized for model evaluation. All images were resized to 224 × 224 pixels, and data augmentation techniques were applied to enhance generalization capability. The ViT model was implemented using TensorFlow and trained with the Adam optimizer (learning rate = 0.0001, batch size = 16, epochs = 50), employing early stopping and learning rate scheduling to mitigate overfitting. The proposed model achieved an overall accuracy of 0.97, with precision, recall, and F1-scores consistently exceeding 0.97. Class-level analysis demonstrated exceptional performance in identifying benign tissue (precision = 0.999, recall = 1.000, F1 = 0.999) and robust classification of malignant subtypes, including adenocarcinoma (F1 = 0.957) and squamous cell carcinoma (F1 = 0.959). These results emphasize the ViT’s strong capability in capturing global contextual features, surpassing conventional CNN-based methods that primarily rely on local feature extraction.