Deep Learning Framework for Automated Lung and Pulmonary Nodule Segmentation in Thoracic CT Scans

Abstract

This thesis presents a deep learning framework for automated segmentation of bilateral lungs and pulmonary nodules from CT scans to support early lung cancer detection. A U-Net model with a ResNet-18 backbone was developed using the LUNA16 dataset, applying Hounsfield unit windowing, data augmentation, and transfer learning to improve performance with limited training data. The model achieved excellent lung segmentation accuracy, with Dice scores of 0.996 for the left lung and 0.973 for the right lung, while pulmonary nodule segmentation achieved a Dice score of 0.848 despite the complexity of small and irregular nodules. Results across multiple independent test sets demonstrated stable and reliable performance. The study highlights the effectiveness of single-pass multi-class segmentation while acknowledging limitations such as limited training data and the use of 2D slice-based processing. Future work will focus on 3D architectures, attention mechanisms, and larger multi-institutional datasets for clinical deployment.