Adaptive alpha-scaled dynamic CBAM integrated into a dynamic attention-guided hybrid deep learning framework for high-precision lung cancer histopathological diagnosis

Lung cancer remains one of the leading causes of death worldwide due to its high mortality rates, which are primarily attributed to the absence of symptoms in the early stages of the disease, making detection challenging. The diagnosis of cancer can be made through the analysis of images obtained from histopathological images, where the classification of the images poses difficulties since it requires dealing with noise, low image contrast, and the complexity of images. Nevertheless, the intrinsic noise, less contrast, and haphazard staining generally present in histopathology images often hinder the performance of CNN-based classification methods. These limitations emphasise the necessity of constructing a strong system that imposes useful approaches for image enhancement methods and monitors the feature extraction process. Thus, the purpose of this research is to formulate a hybrid deep learning approach for lung cancer detection that incorporates the present methods for augmenting image presentation and ensuring adaptive attention to the attributes. For improved performance, hierarchical enhancement of images was adopted to improve the performance of a dynamic convolution block attention mechanism hybrid model aimed at classifying images of cancerous tissues. Using the Lung and Colon Histopathological Image Dataset, the CNN model managed to produce the highest accuracy score of 93.2% when classifying cancer tissue images. After performing the enhancement process, the enhanced CNN models such as Xception, DenseNet121, InceptionV3, MobileNetV2, and others managed to reach an accuracy rate of 95% to 97%. For increased discriminatory learning and interpretability of the model, the dynamic Convolutional Block Attention Module (CBAM) is introduced to modulate spatial and channel attention mechanisms based on the features’ complexity. Through incorporation of the ResNet50 model in image enhancement, it managed to generate an accuracy rate of 99.73%. The results prove the efficacy of the proposed enhancement-attention framework as well as its adoptability for lung cancer diagnostic systems.