Authors : Jaya Shrivastav, Dr. Ranu Pandey
DOI : 10.46335/IJIES.2025.10.4.16
Article Link
Abstract – Lung cancer continues to be one of the deadliest cancers globally, with early detection playing a critical role in improving survival rates. In this work, we present an advanced deep learning framework for lung cancer segmentation using an Attention-Enhanced U-Net architecture with a Hybrid Loss Function. The model was trained and evaluated on a publicly available CT scan dataset from Kaggle. This dataset includes diverse axial CT slices with annotations, providing a suitable foundation for training robust segmentation models. To overcome challenges such as class imbalance and tumor heterogeneity, our approach integrates spatial attention mechanisms into the U-Net architecture and employs a hybrid loss combining Dice and Focal losses. Our proposed model achieves a maximum segmentation accuracy of 92%, with a Dice score of 0.89 and F1-score of 0.83, outperforming traditional U-Net baselines. These results demonstrate the effectiveness of the proposed method in accurately identifying small and irregular tumor regions, making it a promising tool for aiding clinical diagnosis.
