Segmentation of lungs with interstitial lung disease in CT Scans: A TV-L1based texture analysis approach

Abstract

Lung segmentation methods are important for automated lung image analysis tasks such as quantification of lung diseases. In this paper, we describe a method for segmentation of lungs with interstitial lung disease (ILD). In thoracic CT scans, such lungs are characterized by the presence of texture patterns like honeycombing, which makes lung segmentation difficult. We employ a 3D total variation L1 (TV-L1) based texture analysis approach to extract these patterns and attenuate the density of the corresponding voxels in the CT scan. The modified CT scan is then utilized as input to an existing 3D robust active shape model based lung segmentation method. The proposed method was evaluated on 77 CT scans of lungs with and without ILD. On cases with ILD, our method obtained an average volumetric overlap of 0.95±0.02, which was statistically significantly better than two other approaches. The TV-L1 texture analysis utilizes GPUs, making our method fast.