부산대학교 도서관

Lung cancer represents the dominant cause of cancer associated deaths worldwide. Researchers have proposed many methods for medical image analysis of CT scans through the decades but in recent times, deep learning (DL) has gained a lot of consideration in the field of analysis of biomedical images . Image processing techniques are used on image data after extracting CT scan features, to determine if the patient’s found nodule is benign or malignant. The most important of these techniques is segmentation which helps identify the shape,size and volume of the nodule. We experiment with several methods of Convolutional Neural Networks (CNN) to target the best approach for pulmonary nodule segmentation. Our aim is to find the compromise between best accuracy, cheapest processing costs and scalability. First of all, it is important to observe that our method is proposed for the segmentation of lung nodules on a previously defined region of interest (ROI) instead of the whole image as input to our networks to cover the fast processing target and experimenting with 3D U-Net and variations of 3D V-Net. We performed analysis to measures of sensitivity and specificity and compared different segmentation approaches using our proposed modified U-Net with a novel adaptive focal loss function and our implementation of a modified V-Net with different architectures. The modified U-Net with the updated loss gave the best sensitivity and specificity of 0.9132 and 0.9807 respectively.