Deep Learning-Based Lung Cancer Detection from CT scan Using EfficientNetV2 with Spatial Attention and Explainable AI

Abstract

Lung cancer remains the leading cause of cancer-related mortality worldwide, making early and accurate subtype classification critical for timely intervention. This paper presents LungNet256, a deep learning framework for automated lung cancer detection and classification from chest CT images. The model employs an EfficientNetV2-B0 backbone pre-trained on ImageNet, augmented with a custom Spatial Attention Pooling module to emphasize diagnostically relevant regions, classifying scans into four categories: Adenocarcinoma, Large Cell Carcinoma, Squamous Cell Carcinoma, and Normal tissue. A two-stage training strategy—frozen-backbone head training followed by fine-tuning of the last four backbone blocks—combined with Focal Loss and label smoothing addresses class imbalance, while Test-Time Augmentation enhances inference performance. Model interpretability is achieved through EigenCAM and SHAP attributions. On the chest CT dataset, LungNet256 achieves 97.33% test accuracy and a Macro AUC-ROC of 0.9975, highlighting its potential as an effective clinical decision-support tool.