June 2023
Volume 64, Issue 8
Open Access
ARVO Annual Meeting Abstract  |   June 2023
Deep Learning (DL) Based Segmentation of Extraocular Muscles (EOMs) from Orbital Magnetic Resonance Imaging (MRI)
Author Affiliations & Notes
  • Qi Wei
    Bioengineering, George Mason University, Fairfax, Virginia, United States
  • Amad Qureshi
    Bioengineering, George Mason University, Fairfax, Virginia, United States
  • Seongjin Lim
    University of California Los Angeles, Los Angeles, California, United States
  • Soh Youn Suh
    University of California Los Angeles, Los Angeles, California, United States
  • Parag Chitnis
    Bioengineering, George Mason University, Fairfax, Virginia, United States
  • Joseph L Demer
    University of California Los Angeles, Los Angeles, California, United States
  • Footnotes
    Commercial Relationships   Qi Wei None; Amad Qureshi None; Seongjin Lim None; Soh Youn Suh None; Parag Chitnis None; Joseph Demer None
  • Footnotes
    Support  NIH R01EY029715 and EY008313
Investigative Ophthalmology & Visual Science June 2023, Vol.64, 1111. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Qi Wei, Amad Qureshi, Seongjin Lim, Soh Youn Suh, Parag Chitnis, Joseph L Demer; Deep Learning (DL) Based Segmentation of Extraocular Muscles (EOMs) from Orbital Magnetic Resonance Imaging (MRI). Invest. Ophthalmol. Vis. Sci. 2023;64(8):1111.

      Download citation file:


      © ARVO (1962-2015); The Authors (2016-present)

      ×
  • Supplements
Abstract

Purpose : MRI provides excellent visualization of EOMs, clinically valuable and for research on ocular biomechanics. However, manual anatomical segmentation is not only labor intensive but subject to human errors. DL has shown promise in automatic anatomical segmentation. To examine feasibility of DL to objectively analyze MRI morphology of EOMs, we compared several DL models for segmentation and morphometric accuracy.

Methods : MRI obtained in quasi-coronal, 2mm thick planes (resolution 312 mm) using surface coils in target-controlled central gaze was analyzed for both eyes of 39 subjects, providing 78 coronal MR image stacks. In 988 images, the lateral (LR), medial (MR), superior (SR), and inferior rectus (IR) and the superior oblique (SO) EOMs were traced digitally in Fiji to obtain masks of EOM pixels used as ground truth. Performances of four DL models were compared: UNet, UNeXT, DeepLabV3, and ConResNet. 85% (64 eyes, 837 images) were used in training, and 15% (12 eyes, 151 images) were used in testing trained models. Evaluation metrics included Intersection-Over-Union (IoU) and Dice Coefficient, both popular to measure correctly segmented area. We also calculated distance error of the EOM centroid, important to characterize EOM paths.

Results : All DL models could accurately segment most testing images (Fig. A-C), with estimated centroids reasonably close to expected locations (Fig. D), but model performances varied (Figs. E-G). Mean IoU (±SD) of all EOMs in all testing images was 0.77±0.21, 0.73±0.23, 0.73±0.21, and 0.65±0.30, for UNet, UNeXT, DeepLabV3, and ConResNet, respectively. Mean Dice coefficients were 0.84±0.20, 0.81±0.22, 0.82±0.20, and 0.73±0.32, respectively. IoU and Dice for UNet were significantly more accurate than UNeXt, DeepLabV3, and ConResNet. UNet, UNeXT, and DeepLabV3 similarly estimated centroids but were more accurate than ConResNet. Segmentation accuracy varied among different EOMs and among anterior-posterior image planes, being greatest in the mid-orbit (Fig. H).

Conclusions : DL models such as UNet, UNeXt and DeepLabV3 are promising computational methods to automatically extract anatomical descriptions of EOMs from MRI, especially near the EOM bellies. Further research is needed to improve segmentation accuracy near the origins and insertions.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

×
×

This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.

×