Content‐oriented sparse representation (COSR) for CT denoising with preservation of texture and edge

• Published in: Medical physics (Lancaster) Vol. 45; no. 11; pp. 4942 - 4954
• Main Authors: Xie, Huiqiao, Niu, Tianye, Tang, Shaojie, Yang, Xiaofeng, Kadom, Nadja, Tang, Xiangyang
• Format: Journal Article
• Language: English
• Published: United States 01.11.2018
• Subjects: Algorithms; CBCT; denoising; dictionary learning; edge preservation; Image Processing, Computer-Assisted - methods; MDCT; micro‐CT; Signal-To-Noise Ratio; sparse coding; sparse representation; texture preservation; Tomography, X-Ray Computed; denoising; CT; texture preservation; edge preservation; dictionary learning; MDCT; micro-CT; sparse coding; CBCT; sparse representation
• ISSN: 0094-2405; 2473-4209; 2473-4209
• DOI: 10.1002/mp.13189

Purpose Denoising has been a challenging research subject in medical imaging, since the suppression of noise conflicts with the preservation of texture and edges. To address this challenge, we develop a content‐oriented sparse representation (COSR) method for denoising in computed tomography (CT). Methods An image is segmented into a number of content areas and each of them consists of similar material. Having been ex‐painted, each content area is sparsely coded using the dictionary learnt from patches extracted from the corresponding content area. By constraining sparsity, noise is suppressed and the final image is formed by aggregating all denoised content areas. The performance of COSR method is examined with images simulated by computer and generated by multidetector row CT (MDCT), cone beam CT (CBCT), and micro‐CT, in which water phantom, anthropomorphic phantom, a human subject, and a small animal are engaged, using the figures of merit, such as standard division (SD), contrast to noise ratio (CNR), and thresholded edge keeping index (EKIth) and structural similarity index (SSIM). In addition, the optimization of performance by parameter tuning is also investigated. Results Quantitatively gauged by metrics of noise, EKIth and SSIM, the performance evaluation shows that the proposed COSR method is effective in denoising (>50% reduction in noise) while it outperforms the conventional sparse representation method in preservation of texture and edge by ~20% (gauged by SSIM). It has also been shown that the COSR method is tolerable to inaccuracy in content area segmentation and variation in dictionary learning. Moreover, the computational efficiency of COSR can be substantially improved using prelearnt dictionaries. Conclusions The COSR method would find its utility in clinical and preclinical applications, such as low‐dose CT, image segmentation, registration, and computer‐aided diagnosis. The proposal of COSR denoising is of innovation and significance in the theory and practice of denoising in medical imaging. A demonstration code package is available at https://github.com/xiehq/COSR.