# Deep learning based optic nerve sheath diameter characterization and structure quantification on transorbital ultrasound images

## TL;DR

This paper introduces a deep learning model that improves the accuracy of measuring optic nerve structures in ultrasound images, aiding in the diagnosis of neurological conditions.

## Contribution

The novel contribution is a deep neural network with shared/specific feature branches and an uncertainty-aware loss function for robust optic nerve segmentation.

## Key findings

- The model achieved a 73.3% Dice score on optic nerve segmentation.
- It outperformed existing methods with an 84.5% AUROC for optic nerve sheath diameter quantification.
- The model shows strong potential for clinical applications in neuro-ophthalmology.

## Abstract

Optic nerve quantification plays a pivotal role as a biomarker in the non-invasive assessment of elevated intracranial pressure and other neuro-ophthalmic conditions. The manual identification of these optic nerve structures is both resource-intensive and time-consuming. The accuracy of optic nerve segmentation in automated methods directly depends on the quality of the ultrasound images. In instances of sub-optimal image quality, applying deep learning-based methodologies emerges as a more effective approach for precise segmentation. In this work, we propose a deep neural network combining the benefits of shared and specific feature extraction branches as well as the uncertainty-aware loss function. Such an uncertainty-aware loss function could enable the model to learn a robust object structure. Experiments on a multi-center publicly available dataset demonstrate the superior performance of our model in optic nerve segmentation and its strong potential of optic nerve sheath diameter quantification. Specifically, our model has achieved 73.3 % Dice score and 84.5% AUROC on the test dataset, outperforming the state-of-the-art models by a large margin.