Automatic magnetic resonance imaging segmentation of the eye based on 3D Active Shape Models

Abstract

Purpose: Delineation of ocular anatomy in 3D imaging involves a big challenge for ophthalmologists, mostly when developing the treatment planning for ocular diseases. Magnetic resonance imaging (MRI) is used today in clinical practice as a complementary source of information, together with fundus imaging or ultrasound, in treatment planning for retinoblastoma in children. We present here a novel 3D statistical Active Shape Model to automatically segment the eye anatomy in the MRI.

Methods: Our data set is composed of 12 healthy eyes from children aged 3.2±1.7 years. Imaging was performed a 3 Tesla Siemens machine, with a 32-channel surface head coil. Enhanced T1-weighted VIBE (TR/TE, 20/3.91 ms) was acquired with isotropic spatial resolution of 0,41mm. The eye model comprises the regions of the lens, the vitreous humor, the sclera and the cornea. Manual delineations were done for all subjects and structures and corrected for all orthogonal planes. Model generation was done by combining all the surfaces extracted from manual delineation, generating a Statistical Shape Model (SSM), and coupling the intensity information to build the Active Shape Model (ASM). The segmentation of a subject was reduced to an optimization problem where the Mahalanobis distance between the ASM and the subject was minimized. We quantitatively validated our segmentation method with a leave-one-out cross validation test, removing the subject under analysis from the model computation. The previous ASM fitting was applied on the subject. The resulting segmentation was assessed by the overlapping measure (Dice Similarity Coefficient, DSC) with the manual segmentation used as the ground truth.

Results: Through the leave-one-out test we obtain an average DSC of 94.37±2.24% for the regions of the Sclera and the cornea, a 93.62±2.06% DSC for the region of the vitreous humor and a 69.09±6,34% DSC for the region of the lens. These results can be considered as a good fitting, particularly in the posterior part of the human eye, where the retina is located. The execution time for the automatic segmentation is in average less than 7 seconds.

Conclusions: We have shown a reliable and fast segmentation tool that provides a significant benefit for the task of delineating the vitreous humor, the sclera, the cornea and the lens. This tool will now be adapted for clinical use, contributing to diagnosis and treatment planning of intraocular tumors.

Keywords: 549 image processing • 703 retinoblastoma • 550 imaging/image analysis: clinical