Multi‐Stage Deep Learning Model for the Detection and Classification of Lung Nodules From CT Scan Images

ABSTRACT

Early detection of pulmonary nodules in computed tomography (CT) scans is crucial for improving lung cancer diagnosis and survival rates. We propose a novel multi‐stage deep learning framework for automated lung nodule detection and classification, addressing limitations in existing methods such as low sensitivity and high false‐positive rates. Our approach comprises three phases: (1) lung region segmentation using a modified pyramid scene parsing network (PSPNet) with enhanced contextual feature extraction, (2) candidate nodule extraction via a redesigned U‐Net architecture optimized for small nodule detection, and (3) classification of nodules as benign or malignant using a tailored ResNet‐50 with improved gradient flow. We trained and evaluated the framework on the LUNA16 dataset, comprising 888 CT scans with 1186 annotated nodules, split into 710 scans for training, 128 for validation, and 50 for testing. Our model achieved an accuracy of 95.25% and an area under the ROC curve (AUC) of 0.97, outperforming state‐of‐the‐art methods by 3.5% in accuracy and reducing false positives by 12%. These results highlight the framework's robustness and precision, offering a reliable solution for early lung nodule detection with significant potential for clinical lung cancer screening applications.