Performance Analysis of Convolution Neural Network with Different Architecture for Lung Segmentation

As lung cancer is one of the significant causes of death, there is a need for the development of algorithms for early detection of these cancers. Early detection of lung cancer helps to provide appropriate treatment and reduce morbidity. Accurate segmentation of the lung is an essential step in every computer-aided diagnosis (CAD) system to provide an accurate lung CT image analysis. This study is focused on the design of the appropriate architecture of the convolution neural network (CNN) using suitable combinations of CNN blocks to improve lung segmentation efficiency. Based on the scientific intuition, three CNN architectures are proposed for effective segmentation of lung parts from CT images. These CNN architectures are varied by the depth of down sampling of images as 32 × 32, 16 × 16 and 8 × 8. The performances of these CNN are obtained as under segmentation or over-segmentation by comparing the segmented lung part with ground truth lung images. This performance analysis shows the segmentation efficiency greatly affected by appropriate selection of downsampling of these images.