Texture versus shape analysis for lung nodule similarity in computed tomography studies
Marwa Muhammad, Daniela Stan Raicu, Jacob David Furst, Ekarin Varutbangkul · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2008
With the aim of reducing the radiologists' subjectivity and the high degree of inter-observer variability, Content-based Image Retrieval (CBIR) systems have been proposed to provide visual comparisons of a given lesion to a collection of similar lesions of known pathology. In this paper, we present the effectiveness of shape features versus texture features for calculating lung nodules' similarity in Computed Tomography (CT) studies. In our study, we used eighty-five cases of thoracic CT data from the Lung Image Database Consortium (LIDC). To encode the shape information, we used the eight most commonly used shape features for pulmonary nodule detection and diagnosis by existent CAD systems. For the texture, we used co-occurrence, Gabor, and Markov features implemented in our previous CBIR work. Our preliminary results give low overall precision results for shape compared to texture, showing that shape features are not effective by themselves at capturing all the information we need to compare the lung nodules.