Effective lung nodule segmentation and classification by employing a SPPUNet model and global context attention-based...

Lung nodules (LNs) detection using computer tomography (CT) images is essential to reduce the mortality of lung cancer (LC). The complex three-dimensional structure of lung CT data and the variation in the forms and appearances of LNs make the accurate identification of pulmonary nodules still extremely challenging. Although deep learning (DL) methods outperform handcrafted approaches, several challenges remain to be solved. Detecting malignant regions alone is insufficient for clinical decision-making; segmentation by severity and grading analysis are necessary to reduce false positives. Training DL models also requires many datasets, which is difficult in the medical domain due to ethical concerns, limited expert annotations, and the scarcity of disease-specific images. Moreover, insufficient data and class imbalance often lead to overfitting and a reduction in performance. To address these limitations, this study proposes a hybrid pre-trained architecture for more accurate automated pulmonary nodule segmentation and classification. The research comprises 3 phases: preprocessing, segmentation, and classification. Initially, the CT lung images are gathered from the openly available LUNA16 dataset. Then, preprocessing is performed on the collected data by noise filtering using a guided filter and data augmentation to balance the dataset. Doing so improves the data quality for learning and classification processes and prevents the model from biased outcomes. Afterward, the Spatial Pyramid Pooling centered U-shaped network (SPPUNet) is employed to segment the lung regions, enabling the classifier to easily identify and analyze nodules, lesions, and other abnormalities. Finally, the classification is performed using the Global Context Attention integrated InceptionV3 (GCAINCPV3) network, which enables medical professionals to determine the nature of the nodule and provide the most appropriate treatment plan for patients. The outcomes demonstrated that the proposed system outperforms existing systems, achieving an accuracy of 99.23%.