Authors:
Angelo Antonio Manzatto
;
Edson Emílio Justino
and
Edson José Rodrigues Scalabrin
Affiliation:
Programa de Pós Graduação em Informática, PPGIa - PUCPR, Curitiba, Paraná, Brazil
Keyword(s):
Deep Learning, Medical Segmentation, Computed Tomography, Bones.
Abstract:
The segmentation of human body organs in medical imaging is a widely used process to detect and diagnose diseases in medicine and to help students learn human anatomy in education. Despite its significance, segmentation is time consuming and costly because it requires experts in the field, time, and the requisite tools. Following the advances in artificial intelligence, deep learning networks were employed in this study to segment computerized tomography images of the full human body, made available by the Visible Human Project (VHP), which included among 19 classes (18 types of bones and background): cranium, mandible, clavicle, scapula, humerus, radius, ulna, hands, ribs, sternum, vertebrae, sacrum, hips, femur, patella, tibia, fibula, and feet. For the proposed methodology, a VHP male body tomographic base containing 1865 images in addition to the 20 IRCAD tomographic bases containing 2823 samples were used to train deep learning networks of various architectures. Segmentation was
tested on the VHP female body base containing 1730 images. Our quantitative evaluation of the results with respect to the overall average Dice coefficient was 0.5673 among the selected network topologies. Subsequent statistical tests demonstrated the superiority of the U-Net network over the other architectures, with an average Dice of 0.6854.
(More)