Automatic segmentation of the optic nerve in transorbital ultrasound images using a deep learning approach

Transorbital sonography is able to provide reliable information about (a) intra-cranial pressure estimation through the optic nerve sheath diameter (ONSD) measurement, and (b) optic nerve atrophy in patients with multiple sclerosis through the optic nerve diameter (OND). In this study, we present the first method for the automatic measurement of the OND and ONSD using a deep learning technique (UNet with ResNet50 encoder) for the optic nerve segmentation. The dataset included 201 images from 50 patients. The automated measurements were compared with manual ones obtained by one operator. The mean error was equal to 0.07 ± 0.34 mm and -0.07 ± 0.67 mm, for the OND and ONSD, respectively. The developed system should aid in standardizing OND and ONSD measurements and reduce manual evaluation variability.