Positron emission tomography--computed tomography (PET-CT) produces co-registered anatomical (CT) and functional (PET) patient information (3D image set) from a single scanning session, and is now accepted as the best imaging technique to accurately stage the most common form of primary lung cancer--non-small cell lung cancer (NSCLC). This paper presents a content-based image retrieval (CBIR) method for retrieving similar images as a reference dataset to potentially aid the physicians in PET-CT scan interpretation. We design a spatial distribution to describe the spatial information of each region-of-interest (ROI), and a pairwise ROI mapping scheme between images to compute the image matching level. Similar images are then retrieved based on the local and spatial information of the detected ROIs, and a learned weighted sum of ROI distances. Our evaluation on clinical data shows good image retrieval performance.