DEVELOPMENT OF AN AUTOMATED SEGMENTATION ALGORITHM TO IDENTIFY BONES OF THE HAND

Background The evaluation of structural damage with plain radiography is important to clinicians and patients. Standard scoring methods include the Sharp-van der Heijde (SVdH) and Ratingen methods [1] however these systems are time-consuming. Therefore, it is difficult to perform large cohort studies. We set out to develop an automated algorithm to identify bones on plain radiographs as a step towards developing automated quantification of structural damage for use on large datasets.Objectives To develop a novel algorithm to segment outlines of finger bones in hand radiographs.Methods 101 hand radiographs were gathered from the Bath longitudinal cohort (UK). All patients fulfilled the CASPAR criteria for Psoriatic Arthiritis (PSA). None of the patients had damage on SVdH and Ratingen scoring (blinded). The metacarpal (MC), proximal phalanx (PP), middle phalanx (MP), and distal phalanx (DP) in the right index finger were delineated by a rheumatologist. These outlines were used to build a statistical model of the shape using a Gaussian Process Latent Variable Model (GPLVM) [2]. Bones are segmented by matching the shape on a radiograph to the statistical model.Results The performance of the matching algorithm was compared with a traditional algorithm (snakes) using the Adjusted Rand Score (ARND). The ARND score measures the similarity of the segmentation with the ground truth. A perfect segmentation has a score close to 1. We tested the algorithm on 9 PP, 9 MP and 8 DP and 6 MC bones in the right index finger. The results are reported in table 1. We report a mean improvement in ARAND of 0.19, 0.87, 0.43 and 0.30 for the PP, MP, DP and MC respectively.Conclusion We report a reliable algorithm for the identification of metacarpal, proximal, middle and distal phalanx bones of the hand. Future work will focus on using the output of the segmentation algorithm to track damage progression over time.References [1] Van der Heijde, D., PAULUS, H., and SHEKELLE, P. (2000). How to read radiographs according to the Sharp/vander Heijde method. Discussion: Heterogeneity in rheumatoid arthritis radiographic trials. Issues to consider in a metaanalysis. Journal of rheumatology, 27(1):261–263.[2] Titsias, M. K. andLawrence, N. D. (2010). Bayesian Gaussian process latent variable model. In International Conference onArtificial Intelligence and Statistics, pages 844–851.View this table:Table 1 Adjusted RAND scores for comparing our algorithm to a traditional one (snakes)Key: Adjusted Rand Score (ARND) score measures the similarity of the segmentation with the ground truth. A perfect segmentation has a score close to 1. Metacarpal (MC), proximal phalanx (PP), middle phalanx (MP), and distal phalanx (DP)Figure 1 Shape matching algorithm output demonstrating segmented outlines of the DP, MP, PP and MC in red, green, orange, and blue respectively.Disclosure of Interests Adwaye Rambojun: None declared, William Tillett Grant/research support from: AbbVie, Celgene, and Lilly, Consultant for: AbbVie, Celgene, Lilly, Novartis, and Pfizer, Speakers bureau: Abbvie, Celgene, Lilly, Janssen, Novartis, UCB, and Pfizer, Neill Campbell: None declared, Tony Shardlow: None declared