Comprehensive and Comparative Global and Local Feature Extraction Framework for Lung Cancer Detection Using CT Scan Images

• Published in: IEEE access 2021, Vol.9, p.158140-158154
• Main Authors: Alzubaidi, Mohammad A., Otoom, Mwaffaq, Jaradat, Hamza
• Format: Article
• Language: English
• Subjects: Accuracy; Algorithms; Cancer; Computed tomography; CT scan; Data collection; Feature extraction; Gabor filter; Gabor filters; global feature extraction; Histograms; Image warping; local feature extraction; Lung; Lung cancer; Machine learning; Medical imaging; Sensitivity; Support vector machines; SVM; Training; Tumors
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2021.3129597

Lung cancer is reported to be the second most common cancer disease. This paper proposes a comprehensive and comparative global and local feature extraction framework for lung cancer detection using CT scan images. This framework consists of three main phases: data collection, global training and testing, and local training and testing. A set of 1000 CT scan images is used in this study. During the global training and testing phase, the collected images are preprocessed through image warping and cropping. Global features are then extracted from images to represent each image with feature vectors, using ten different image feature types. The feature vectors are then used to build detection models with six different machine learning algorithms. In the local training and testing phase, each image is divided into a set of local blocks. Those feature types that performed well in the global phase are then extracted from each of these blocks, to represent each block with feature vectors. These feature vectors are then used to build detection models for all of the image blocks, using the learning algorithms that performed well in the global phase. The results show that the Gabor Filter, the Histogram of Oriented Gradients (HOG), and the Haar Wavelet feature types outperformed the other seven feature types. The results also show that Support Vector Machine (SVM) outperforms the other five learning algorithms. Of most importance, the proposed local feature extraction approach outperforms the traditional global one. In the local phase, using SVM with Haar Wavelet features achieved 90% accuracy, 88% sensitivity, and 91% specificity. Using SVM with HOG features achieved 88% accuracy, 85% sensitivity, and 89% specificity. Finally, using SVM with Gabor Filter features achieved the best accuracy, sensitivity, and specificity rates of 97%, 96%, and 97%, respectively.