Authors - Mamatha Kurra, Ochin Sharma, G S Pradeep Ghantasala Abstract - Accurate segmentation of pulmonary nodules in low-dose CT (LDCT) scans plays a crucial role in the early detection of lung cancer. However, small and irregular nodules remain difficult to detect due to low contrast, anatomical variability, and imaging artifacts. In this study, we perform a comparative evaluation of widely used deep learning-based segmentation architectures-namely, vanilla U-Net, Feature Pyramid Network (FPN), and Mask R-CNN-on benchmark datasets LIDC-IDRI and LUNA16. Building on the observed limitations of these models, we introduce a refined Hybrid U-Net architecture augmented with attention gates and Squeeze-and-Excitation (SE) blocks. This enhancement improves the model’s ability to focus on clinically relevant features while maintaining strong spatial consistency across encoder-decoder layers. Preprocessing involves Hounsfield Unit (HU) windowing (−1000 to 400 HU) to isolate lung parenchyma, followed by patch extraction (128×128) to better represent small nodules and manage class imbalance. The model is trained using a compound loss function that combines Dice loss and Boundary loss in a 0.7:0.3 ratio to balance volumetric overlap and edge accuracy. Experimental results on the LIDC-IDRI dataset show that the proposed attention guided model achieves a Dice coefficient exceeding 0.85, outperforming the baseline U-Net (average Dice 0.78). Evaluation metrics such as sensitivity (true positive rate) further confirm the effectiveness of our approach in capturing subtle nodule features. This work demonstrates that integrating attention mechanisms and feature recalibration into U-Net significantly boosts segmentation performance on challenging medical imaging tasks. Our results provide a strong foundation for deploying more accurate and interpretable tools in computer-aided diagnosis pipelines for lung cancer screening.
Open Zoom