Métodos para segmentação de medula espinhal e esôfago em tomografia computadorizada de planejamento à radioterapia

Abstract

Radiotherapy (RT) is one of the most used options for the treatment of cancer. For planning treatments with RT, computed tomography (CT) exams are commonly used to define organs at risk (OARs) and analyze the impact of radiation on these. Delimiting these OARs takes a lot of time from medical specialists, and it involves a large team of professionals. Besides, this task performed slice by slice becomes exhausting and, consequently, subject to errors, especially in organs such as the spinal cord and esophagus, which span several CT slices and require precise segmentation. Thus, in this work, computational methods capable of segmenting the spinal cord and esophagus in planning CT are proposed. For spinal cord segmentation, two methods are proposed, the first being comprised of techniques such as adaptive template matching for initial segmentation, a superpixel technique known as intrinsic manifold simple linear iterative clustering (IMSLIC) for segmentation of spinal cord candidates and convolutional neural networks (CNN) to classify these candidates. The second method for spinal cord segmentation consists of techniques, such as adaptive template matching for initial segmentation of a region of interest and fully convolutional neural networks (FCN) for final spinal cord segmentation. In the esophagus segmentation, it is proposed a method composed of techniques to registration the volume of patients, atlas to define the volume of interest where the esophagus is located, pre-processing to highlight the walls of the esophagus, CNN for esophagus segmentation, and post-processing to refine segmentation. We applied the method to 36 planning CT images provided by The Cancer Imaging Archive. The first method for spinal cord segmentation obtained an accuracy of 92.55%, a specificity of 92.87%, a sensitivity of 89.23%, and 78.20% of Dice. The second method for spinal cord segmentation obtained an accuracy of 99.24%, a specificity of 99.38%, a sensitivity of 91.52%, and 81.69% of Dice. The esophagus segmentation method obtained an accuracy of 99.69%, a specificity of 99.76%, a sensitivity of 90.61%, and 82.15% of Dice. Thus, we demostrated the feasibility of analyzing CT planning using image processing and pattern recognition techniques.