Spatial Attention-Enhanced Deep Learning for Lung and Colon Cancer Histopathological Subtype Classification

Abstract

The process of diagnosing cancer involves detecting abnormal cell growth through clinical examination and imaging techniques. Effective cancer treatment is based on the classification of lung and colon cancer types based on the histopathological images; however, analyzing these images manually takes a lot of time and frequently leads to inconsistent detections. This study proposes a deep learning- based approach incorporating a spatial attention mechanism to enhance cancer subtype classification. The pretrained ResNet50 architecture is applied on three convolutional neural network models: single task, multitask, and spatial attention models for comparison. The LC25000 dataset with 25,000 histopathological images of five classes of tissues were used for the experimentation process. Performance evaluation metrics like accuracy, precision, recall and macro F1-score were used to evaluate the models. The results demonstrate that attention-based models outperform others, achieving an accuracy of 99.82% and a macro F1-score of 0.9982. Furthermore, attention heatmaps help in concentrating on the critical morphological features, like dense nuclei. These findings indicate that integrating spatial attention mechanisms significantly enhances the discriminative capability of deep learning models for histopathological cancer classification, supporting their potential for clinical application.