Computer aided detection system for early cancerous pulmonary nodules by optimizing deep learning features

In this paper, a deep learning technique for the early detection of pulmonary nodules from low dose CT (LDCT) images is proposed. The proposed technique is composed from four stages. Firstly, a preprocessing stage is applied to enhance image contrast of low dose images. Secondly, a transfer learning is utilized to extract deep learning features that describe the LDCT images. Thirdly, a genetic algorithm (GA) is learned on the extracted deep learning features using a training subset of the data to optimize the feature-set and select the most relevant features for cancerous nodules detection. Finally, a classification step of the selected features is performed using supported vector machines (SVM) to detect cancerous pulmonary nodules. Preliminary results on a number of 320 LDCT images acquired from 50 different subjects from the International Early Lung Cancer Action Project, I-ELCAP, online public lung image database has achieved a detection accuracy of 92.5%, sensitivity of 90%, and specificity of 95% Comparison results has shown the outstanding results of the proposed method. These preliminary results confirm the promising of our proposed method.