Automatic segmentation and volume measurement of anterior visual pathway in brain 3D-T1WI using deep learning

Yongliang Han ¹, Haixiang Wang ², Qi Luo ¹

⁰Department of Radiology, The First Affiliated Hospital of Chongqing Medical University, Chongqing, China.

Frontiers in Medicine +

|

May 13, 2025

View abstract on PubMed

Summary

The 3D UX-Net deep learning model accurately segments the anterior visual pathway (AVP) in brain T1WI, providing reliable volume measurements for clinical use.

Area Of Science

Medical Imaging
Artificial Intelligence
Neuroscience

Background

Accurate segmentation of the anterior visual pathway (AVP) is crucial for clinical applications.
Manual AVP delineation is time-consuming and resource-intensive, necessitating automated solutions.

Purpose Of The Study

To evaluate the feasibility of automatic AVP segmentation and volume measurement using brain T1-weighted imaging (T1WI).
To assess the performance of the 3D UX-Net deep-learning model for AVP segmentation and compare it with other state-of-the-art models.

Main Methods

Retrospective analysis of brain T1WI from 119 adults, with manual AVP annotation by two radiologists.
Training and validation of a 3D UX-Net model, with performance evaluation using Dice Similarity Coefficient (DSC), 95% Hausdorff Distance (HD95), and Average Symmetric Surface Distance (ASSD).
Comparison of 3D UX-Net against 3D U-Net, UNETR, UNETR++, and Swin SMT models for segmentation accuracy and volume measurement.

Main Results

3D UX-Net achieved the highest DSC (0.893 ± 0.017) and lowest median ASSD (0.234 mm).
Automatic segmentation yielded a mean AVP volume of 1446.78 ± 245.62 mm³, closely matching manual segmentations (VD = 0.068 ± 0.064).
Significant sex-based AVP volume differences were observed (p < 0.001), with no age correlation.

Conclusions

The 3D UX-Net model demonstrates high accuracy for automatic AVP segmentation and volume measurement in brain T1WI.
This study provides normative values for automatic MRI measurement of the AVP in adults.
The developed model offers a time-efficient and reliable alternative to manual AVP delineation in clinical practice.

Keywords:

convolutional neural network deep learning magnetic resonance imaging medical image processing optic nerve