An Explainable Hybrid Transformer-Based Lightweight Framework for Lung Cancer Classification

Lung cancer is one of the deadliest causes of cancer related diseases which actually demands more high accurate diagnosis. The way histopathological analysis is diagnosing around the world is time consuming, also not that much of accurate detecting the cancer, on the other hand Convolutional Neural Networks (CNNs) struggling so hard to cope with its long-range dependencies. For this instance, Vision Transformers (ViTs) are doing very well in terms of computational operations. MBv3-ViT, a lightweight hybrid architecture has been introduced in this paper for lung cancer classification from histopathological image. This model implements the computational efficiency of MobileNetV3 for especially robust local feature extraction with the help of global reasoning of a ViTs. The MobileNetV3 generates a feature with a 7x7x576 map, which is also projected to a 768-dimensional token sequence. Then it is processed through an 8-block ViT encoder. 15,000 images from LC25000 have been trained by us. We have trained encompassing benign, adenocarcinoma also the squamous cell carcinoma tissues. 99.80% accuracy and a 99.80% F1-Score has been achieved by the proposed MBv3-ViT model which also demonstrated the excellent performance. On the other hand, the benchmark models like LEViT, CoAtNet, and HorNet are significantly outperforming. Also, Grade-CAM has been integrated for Explainable AI(XAI), increasing the model trustworthiness. For assisting the pathologist in clinical practices, the high accuracy, lightweight design, and interpretability of MBv3-ViT creates its potential as a valuable and useful tool.