A deep learning method with residual blocks for automatic spinal cord segmentation in planning CT. (January 2022)

• Record Type: Journal Article
• Authors: Bandeira Diniz, João Otàvio; Ferreira, Jonnison Lima; Bandeira Diniz, Pedro Henrique; Silva, Aristófanes Corrêa; Paiva, Anselmo Cardoso
• Is Part Of: Biomedical signal processing and control. Volume 71(2022)Part A
• Publication Date: 2022-01
• Subjects: Computer-aided detection -- Medical images -- Planning CT -- Radiotherapy -- Spinal cord -- U-Net; 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.2021.103074
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: This work investigates a method for automatic spinal cord segmentation in planning CT. The method is composed of template matching and deep learning techniques. The method was applied in 36 CT scans with an average of 200 slices. We proposed a novel deep convolutional neural network. The method achieved an accuracy of 99.35% and a Dice index of 85.47%. Abstract: Radiotherapy (RT) is one of the most used types of cancer treatment. One of its steps is to build a three-dimensional model of the patient's body, usually based on computerized tomography (CT). This model is used to locate the target tissues and their surrounding organs that could be affected by mistake. These organs, called organs at risk (OARs), must be protected from radiation. To accurately define the area where the radiation will be applied and protect the OARs, the specialist must delineate them through manual segmentation. However, in large organs such as the spinal cord that comprises almost all slices of CT, this task can be time-consuming and exhaustive and, therefore, susceptible to errors. Motivated by the problem of manual segmentation and the difficulty that this process brings to specialists, this paper presents a method for automatic spinal cord segmentation in planning CT for radiotherapy. The proposed method is mainly composed of a template matching technique and a novel deep convolutional neural network with residual blocks. To evaluate its performance, it was applied in a CT database of 36 patients. The best model achieved an accuracy of 99.35%, a specificity of 99.57%, a sensitivity of 91.52%, and a Dice index of 85.47%, without any segmentation refinement techniques.