Similarity clustering‐based atlas selection for pelvic CT image segmentation

• Published in: Medical physics (Lancaster) 2019-05, Vol.46 (5), p.2243-2250
• Main Authors: Kennedy, Angel, Dowling, Jason, Greer, Peter B, Holloway, Lois, Jameson, Michael G, Roach, Dale, Ghose, Soumya, Rivest‐Hénault, David, Marcello, Marco, Ebert, Martin A
• Format: Article
• Language: English
• Subjects: autosegmentation; clustering; image registration; image‐atlas
• ISSN: 0094-2405; 2473-4209
• DOI: 10.1002/mp.13494

Purpose To demonstrate selection of a small representative subset of images from a pool of images comprising a potential atlas (PA) pelvic CT set to be used for autosegmentation of a separate target image set. The aim is to balance the need for the atlas set to represent anatomical diversity with the need to minimize resources required to create a high quality atlas set (such as multiobserver delineation), while retaining access to additional information available for the PA image set. Methods Preprocessing was performed for image standardization, followed by image registration. Clustering was used to select the subset that provided the best coverage of a target dataset as measured by postregistration image intensity similarities. Tests for clustering robustness were performed including repeated clustering runs using different starting seeds and clustering repeatedly using 90% of the target dataset chosen randomly. Comparisons of coverage of a target set (comprising 711 pelvic CT images) were made for atlas sets of five images (chosen from a PA set of 39 pelvic CT and MR images) (a) at random (averaged over 50 random atlas selections), (b) based solely on image similarities within the PA set (representing prospective atlas development), (c) based on similarities within the PA set and between the PA and target dataset (representing retrospective atlas development). Comparisons were also made to coverage provided by the entire PA set of 39 images. Results Exemplar selection was highly robust with exemplar selection results being unaffected by choice of starting seed with very occasional change to one of the exemplar choices when the target set was reduced. Coverage of the target set, as measured by best normalized cross‐correlation similarity of target images to any exemplar image, provided by five well‐selected atlas images (mean = 0.6497) was more similar to coverage provided by the entire PA set (mean = 0.6658) than randomly chosen atlas subsets (mean = 0.5977). This was true both of the mean values and the shape of the distributions. Retrospective selection of atlases (mean = 0.6497) provided a very small improvement over prospective atlas selection (mean = 0.6431). All differences were significant (P