Transfer learning-based classification model for the Computed Tomography scan pulmonary images
Vidhi Bishnoi,Nidhi Goel
DOI: https://doi.org/10.1007/s11042-024-19098-8
IF: 2.577
2024-04-13
Multimedia Tools and Applications
Abstract:In recent years, transfer learning has emerged as the most effective method for detecting and classifying lung cancer. Early-stage lung cancer diagnosis using multiple slices of computed tomography scan lung images is challenging. The manual diagnosis of such images is tedious, time-consuming, and biased. Several automated diagnosis systems have been widely used to process computed tomography scan lung images. However, existing diagnostic systems are dependent on annotated datasets. To address the stated problems, a robust automated diagnosis solution has been proposed to classify computed tomography scan lung images without relying on annotations. The proposed model utilizes a transfer learning-based architecture to classify segmented lung images. Segmentation of CT scan lung images is performed using K-means clustering with parameter tuning followed by morphological operations. The number of clusters has been chosen based on the value of the silhouette score. The best silhouette score of 0.66 was obtained during analysis for k=2 clusters. The proposed model achieved a test accuracy; 0.926, a precision; 0.932, recall and F1 score; 0.926, a kappa score; 0.848, and an AUC of 0.904 on the LIDC dataset. For the NLST dataset, it achieved test results; 0.970 accuracy, precision, recall, 0.956 F1 score, 0.934 kappa value, and 0.978 AUC. In addition, the proposed method outperforms state-of-art models with an accuracy of 1-3%, recall of 2-3%, a precision of 2-3%, and an F1 score of 2-4% for the classification of lung cancer. These results justify that the proposed model improves the performance of diagnosis of lung images. Overall performance improvement, robustness in handling various sizes and shapes of lung images, and use of the silhouette score to choose the number of clusters for segmentation make the proposed approach distinct from the existing techniques.
computer science, information systems, theory & methods,engineering, electrical & electronic, software engineering