Abstract

Early detection of lung cancer significantly improves patient survival, with pulmonary nodules in computed tomography (CT) imaging serving as key indicators of early-stage disease. However, manual classification of nodules is time-consuming and subject to interobserver variability. This thesis proposes a reproducible deep learning framework for pulmonary nodule malignancy classification using volumetric CT data. Using the LIDC-IDRI dataset, a standardized preprocessing pipeline is developed, including isotropic resampling, intensity normalization, and 3D patch extraction. A baseline 3D convolutional neural network (CNN) is implemented and compared against a proposed SE-ResNet3D architecture that incorporates residual connections and squeeze-and-excitation mechanisms. Both models are trained and evaluated under identical conditions using patient-level 10-fold cross-validation to ensure fairness and reproducibility. Experimental results demonstrate that the proposed architecture outperforms the baseline model across multiple evaluation metrics. In particular, the SE-ResNet3D achieves a higher ROC-AUC (0.865 vs. 0.742), along with improvements in accuracy, precision, specificity, and F1-score, while maintaining comparable sensitivity. The reduction in false positives indicates improved prediction reliability, which is critical in medical decision-making contexts. These findings suggest that integrating residual learning and channel attention mechanisms enhances the ability of 3D CNNs to capture discriminative features in volumetric medical imaging data. The proposed framework contributes a reproducible pipeline and an improved model for pulmonary nodule malignancy classification, with potential to support clinical decisionmaking and future research in medical image analysis.

Publication Date

5-2026

Document Type

Thesis

Student Type

Graduate

Degree Name

Professional Studies (MS)

Department, Program, or Center

Graduate Programs & Research

Advisor

Ioannis Karamitsos

Comments

This thesis has been embargoed. The full-text will be available on or around 5/1/2027.

Campus

RIT Dubai

Download

Share

COinS