Structure-Adaptive Feature Extraction and Representation for Multi-modality Lung Images Retrieval

Content-based image retrieval (CBIR) has been an active research area since mid 90's with major focus on feature extraction, due to its significant impact on image retrieval performance. When applying CBIR in the medical domain, different imaging modalities and anatomical regions require different feature extraction methods that integrate some domain-specific knowledge for effective image retrieval. This paper presents some new CBIR techniques for positron emission tomography - computed tomography (PET-CT) lung images, which exhibit special characteristics such as similar image intensities of lung tumors and soft tissues. Adaptive texture feature extraction and structural signature representation are proposed, and implemented based on our recently developed CBIR framework. Evaluation of the method on clinical data from lung cancer patients with various disease stages demonstrates its benefits.