LungSightNet: Deep Learning Driven Lung Cancer Prediction Using Compact Convolutional Transformers

Abstract

Accurate and interpretable medical image classification remains a critical challenge in computer-aided diagnosis, particularly under limited dataset conditions. Deep learning models often struggle to capture both local and global patterns simultaneously, and their black-box nature limits clinical trust. Incorporating attention mechanisms and explain- ability techniques can enhance both performance and interpretability. This research proposes a modified Self-Attention Compact Convolutional Transformer (SA-CCT) architecture designed to improve feature extraction and classification performance for lung cancer detection. The model integrates enhanced self-attention mechanisms and custom MLP blocks within the transformer framework, coupled with an improved patch based tokenization strategy that captures richer local and global features from grayscale CT images. To ensure interpretability, Grad-CAM explainability and segmentation based visualization modules are incorporated that enables spatial localization of discriminative regions and validation of model focus. The proposed SA-CCT model is trained and evaluated on a comprehensive lung cancer dataset that achieves 99% of classification accuracy, along with robust performance across per-class metrics and confusion analysis. These results show that the modified SA-CCT architecture is a very successful and easy to understand way to automatically diagnose lung cancer. It also provides a framework that may be used again and again for future research in medical image analysis and transformer-based categorization. Keywords: Lung Cancer Classification, CT Image, SA-CCT, CNN-Transformer Hybrid, Self-Attention, Explainable AI, Grad-CAM, Convolutional Tokenization, Custom MLP Block, Sequence Pooling.