Target organ non-rigid registration on abdominal CT images via deep-learning based detection. (September 2021)

• Record Type: Journal Article
• Authors: Yang, Shao-di; Zhao, Yu-qian; Yang, Zhen; Wang, Yan-jin; Zhang, Fan; Yu, Ling-li; Wen, Xiao-bin
• Is Part Of: Biomedical signal processing and control. Volume 70(2021)
• Publication Date: 2021-09
• Subjects: Non-rigid registration -- Deep learning -- ROI detection -- Abdominal CT image; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.102976
• Languages: English
• ISSNs: 1746-8094

Abstract

Graphical abstract: Highlights: An automatic ROI-based abdominal CT registration framework is proposed. A DCNN-based abdominal CT multi-organ detection model is proposed. A robust intensity-based method is developed to register target organ ROIs. It outperforms some existing methods on abdominal CT organs registration. Abstract: Abdominal target organ registration is essentially important for medical diagnosis and clinical treatment, and the main challenge comes from the complex non-rigid deformations of organ structure and volume. In this paper, an improved deep convolutional neural network (DCNN) is first proposed to automatically detect the abdominal computed tomography (CT) target organ regions of interest (ROIs), in which a CoordConv layer is added to obtain more coordinate information for different target organs, and a transfer learning technique is used for pre-training on a non-medical database to deal with the medical training data scarce situation. Then, the pair-wise target organ ROIs are registered by an intensity-based dissimilarity measure combined with a standard Tikhonov regularization. Finally, the proposed method is evaluated on a self-created clinical database and several public databases. The experimental results demonstrate that our method achieves high accuracy on the target organ ROIs detection with the mean intersection over union (mIOU) and mean average precision (mAP) values of 0.8719 and 0.9445, respectively, and the registration performance on these ROIs is superior to that of some existing methods. Moreover, the proposed method is capable of detecting and registering abdominal target organ ROIs on a single GPU with good network convergence and less time consumption.