Date of Award
Fall 12-4-2025
Document Type
Thesis
Publication Status
Version of Record
Submission Date
December 2025
Department
Physics
College Granting Degree
Charles E. Schmidt College of Science
Department Granting Degree
Physics
Degree Name
Master of Science (MS)
Thesis/Dissertation Advisor [Chair]
Wazir Muhammad
Abstract
Lung cancer is still one of the predominant reason of cancer deaths around the world. It usually starts with small lung nodules that can be seen on CT scans. It is very important to find these nodules early and correctly classify them so that the diagnosis and treatment can be done quickly and effectively. Nevertheless, it is difficult to differentiate between benign and malignant nodules because of their comparable imaging appearance and the variation in manual interpretation among radiologists. To overcome these constraints, our study introduces a deep learning-based CNN model for the automated classification of lung nodules in CT images, aiming to facilitate early detection and alleviate the workload of radiologists. A total of 130 patient cases with corresponding XML annotations were obtained from the LIDC-IDRI dataset. Patient-wise splitting was performed to avoid data overlap, dividing cases into 70% training, 15% validation, and 15% testing sets. Using the first radiologist’s reading, only nodules with malignancy scores of 1-2 (benign) and 4-5 (malignant) were included, while ambiguous cases (score 3) were excluded. Around each valid nodule, 128×128-pixel image patches were extracted, converted to RGB, resized to 224×224 pixels. Manual augmentation (flips and +10° rotation) was applied to benign patches in training and validation sets to address class imbalance, yielding 948 total patches (720 training, 130 validation, and 98 testing). The proposed network was based on EfficientNetB0 (pretrained on ImageNet), serving as a fixed feature extractor, followed by a GlobalAveragePooling2D layer, dropout (0.2), and a final sigmoid-activated dense layer for binary classification. The model was trained on the training data with validation monitoring and evaluated on independent test data at both patch and patient levels. At the patch-level, it achieved 81% accuracy, 87% precision, 60% specificity, 88% recall, and an 87% F1-score. At the patient-level, the model achieved 73% accuracy, 67% precision, 86% recall, 75% F1score, and 63% specificity. Grad-CAM was used for interpretability to highlight the regions that influenced model predictions. Overall, the proposed EfficientB0-based method demonstrates the strong potential for binary lung nodule classification from CT patches. Given the dataset size and encouraging performance outcomes, this approach shows promise for future development of reliable AI-based tools for lung cancer diagnosis.
Recommended Citation
Mahtab, Amen, "DEEP LEARNING-BASED LUNG NODULE CLASSIFICATION USING CT SCANS" (2025). Electronic Theses and Dissertations. 205.
https://digitalcommons.fau.edu/etd_general/205