Minglei Yang1, Hui Ding2, Jingang Kang3, Longfei Cong4, Lei Zhu5, Guangzhi Wang6. 1. Department of Biomedical Engineering, School of Medicine, Tsinghua University, Room C249, Beijing 100084, China. Electronic address: yangml122@gmail.com. 2. Department of Biomedical Engineering, School of Medicine, Tsinghua University, Room C249, Beijing 100084, China. Electronic address: dinghui@tsinghua.edu.cn. 3. Beijing Shen Mindray Medical Electronics Technology Research Institute Co. Ltd, 3 F, Building5, No.8 Chuangye Road, Shangdi Haidian District, Beijing 100085, China. Electronic address: kangjingang@mindray.com. 4. Beijing Shen Mindray Medical Electronics Technology Research Institute Co. Ltd, 3 F, Building5, No.8 Chuangye Road, Shangdi Haidian District, Beijing 100085, China. Electronic address: conglongfei@mindray.com. 5. Beijing Shen Mindray Medical Electronics Technology Research Institute Co. Ltd, 3 F, Building5, No.8 Chuangye Road, Shangdi Haidian District, Beijing 100085, China. Electronic address: zhulei@mindray.com. 6. Department of Biomedical Engineering, School of Medicine, Tsinghua University, Room C249, Beijing 100084, China. Electronic address: wgz-dea@tsinghua.edu.cn.
Abstract
PURPOSE: Ultrasound (US)-magnetic resonance (MR) fusion imaging is a profitable tool for image-guided abdominal diagnosis and biopsy. However, the automatic registration of liver US and MR images remains a challenging task. An effective local structure orientation descriptor (LSOD) for use in registering multimodal images is proposed in this study. METHODS: LSOD utilizes a normalized similarity distance vector of intra-image patch pairs to extract intensity change orientations from intensity value changes in a local area. The multimodal similarity measure is then derived using the LSOD vector difference. Experiments were performed on simulated US and liver 2D US-3D MR images from a phantom, two healthy volunteers, and seven patients. RESULTS: Using the LSOD-based method, the root-mean-square target registration errors (RMS-TREs) were 1.76±1.90mm/2.03±0.84mm in phantom/clinical experiments. All of the results outperformed those obtained using modality independent neighborhood descriptor (MIND)- and linear correlation of linear combination (LC(2))-based methods (phantom/clinical: 5.23±3.35mm/4.32±3.63mm and 9.79±5.03mm/6.29±3.85mm, respectively). The registration cover range for all subjects of the LSOD-based method was 9.16mm, which was larger than those of the MIND- and LC(2)-based methods (5.06 and 5.12mm, respectively). CONCLUSIONS: The results demonstrated that the LSOD-based registration method could robustly register 2D US and 3D MR images of different liver sections with acceptable accuracy for clinical requirements. This approach is useful for the practical clinical application of the US-MR fusion imaging technique.
PURPOSE: Ultrasound (US)-magnetic resonance (MR) fusion imaging is a profitable tool for image-guided abdominal diagnosis and biopsy. However, the automatic registration of liver US and MR images remains a challenging task. An effective local structure orientation descriptor (LSOD) for use in registering multimodal images is proposed in this study. METHODS: LSOD utilizes a normalized similarity distance vector of intra-image patch pairs to extract intensity change orientations from intensity value changes in a local area. The multimodal similarity measure is then derived using the LSOD vector difference. Experiments were performed on simulated US and liver 2D US-3D MR images from a phantom, two healthy volunteers, and seven patients. RESULTS: Using the LSOD-based method, the root-mean-square target registration errors (RMS-TREs) were 1.76±1.90mm/2.03±0.84mm in phantom/clinical experiments. All of the results outperformed those obtained using modality independent neighborhood descriptor (MIND)- and linear correlation of linear combination (LC(2))-based methods (phantom/clinical: 5.23±3.35mm/4.32±3.63mm and 9.79±5.03mm/6.29±3.85mm, respectively). The registration cover range for all subjects of the LSOD-based method was 9.16mm, which was larger than those of the MIND- and LC(2)-based methods (5.06 and 5.12mm, respectively). CONCLUSIONS: The results demonstrated that the LSOD-based registration method could robustly register 2D US and 3D MR images of different liver sections with acceptable accuracy for clinical requirements. This approach is useful for the practical clinical application of the US-MR fusion imaging technique.