Self-adaptive weighted level set evolution based on local intensity difference for parotid ducts segmentation

• Published in: Computers in biology and medicine Vol. 114; p. 103432
• Main Authors: Deng, Xuan, Lan, Tianjun, Chen, Zhifeng, Zhang, Minghui, Tao, Qian, Lu, Zhentai
• Format: Journal Article
• Language: English
• Published: United States Elsevier Ltd 01.11.2019; Elsevier Limited
• Subjects: Algorithms; Computed tomography; Evolution; Image processing; Image segmentation; Impact analysis; Inhomogeneity; Internal Medicine; Level set method; Local mean difference; Local variance difference; Medical diagnosis; Medical imaging; Methods; Metric space; Neighborhoods; Noise; Noise intensity; Other; Parameter sensitivity; Parotid gland; PDs segmentation; Self-adaptive weighted; PDs segmentation; Local variance difference; Local mean difference; Self-adaptive weighted; Level set method
• ISSN: 0010-4825; 1879-0534; 1879-0534
• DOI: 10.1016/j.compbiomed.2019.103432

Parotid ducts (PDs) play an important role in the diagnosis and treatment of parotid lesions. Segmentation of PDs from Cone beam computed tomography (CBCT) images has a significant impact to the pathological analysis of the parotid gland. Although level set methods (LSMs) have achieved considerable success in medical imaging segmentation, it is still a challenging task for existing LSMs to precisely and self-adaptively segment PDs from parotid duct (PD) images with both noise, intensity inhomogeneity, and vague boundary. In this paper, we propose a novel Self-adaptive Weighted level set method via Local intensity Difference (SWLD) to comprehensively solve the above issues. Firstly, a new adaptive weighted operator based on local intensity variance difference has been proposed to overcome the limitations of previous LSMs that are sensitive to parameters, which achieves the aim of automatic segmentation. Secondly, we introduce local intensity mean difference into the energy function to improve the curve evolution efficiency. Thirdly, we eliminate the effects of intensity inhomogeneity, noise, and boundary blur in the parotid image through a local similarity factor with two different neighborhood sizes. Using the same dataset, segmentation of PDs is performed using the proposed SWLD algorithm and existing LSM algorithms. The mean Dice score for the proposed algorithm is 91.3%, and the corresponding mean Hausdorff distance (HD) is 1.746. Experimental results demonstrate that the proposed algorithm is superior to many existing level set segmentation algorithms, and it can accurately and automatically segment the PDs even in complex gradient boundaries. •Proposed method performs better in segmenting parotid duct images with noise, intensity inhomogeneity, and blurred border.•We introduces a new adaptive weighted operator to remove sensitivity to parameters.•Local intensity mean difference accelerates the convergence of weights of local energy terms.