Related Experiment Video
Updated: Oct 6, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.0K
AM-UNet: automated mini 3D end-to-end U-net based network for brain claustrum segmentation
Ahmed Awad Albishri1,2, Syed Jawad Hussain Shah1, Seung Suk Kang3
1School of Computing and Engineering, University of Missouri-Kansas City, Kansas City, MO 64110 USA.
Multimedia Tools and Applications
|January 17, 2022
Summary
We developed AM-UNet, an optimized deep learning model for segmenting the human brain claustrum (CL) from 3D MRI scans. This automated method achieves state-of-the-art accuracy, overcoming segmentation challenges.
Area of Science:
- Medical Image Analysis
- Deep Learning
- Neuroimaging
Background:
- Deep learning (DL) shows promise for medical image segmentation.
- U-Net based models are widely used but under-explored for human brain claustrum (CL) segmentation.
- CL segmentation is difficult due to its structure, data variability, and imperfections.
Purpose of the Study:
- To propose an automated, optimized U-Net based 3D segmentation model for the human brain claustrum.
- To address challenges in CL segmentation including its thin structure and data imbalance.
- To provide a lightweight and scalable solution for CL segmentation.
Main Methods:
- Developed AM-UNet, an end-to-end 3D U-Net model incorporating pre- and post-processing techniques.
- Utilized combined T1/T2 MRI datasets for training and evaluation.
- Performed comparative analysis against existing segmentation models.
Main Results:
- AM-UNet achieved state-of-the-art accuracy for automatic CL segmentation on 3D MRI.
- Achieved Dice score of 82%, Intersection over Union (IoU) of 70%, and Intraclass Correlation Coefficient (ICC) of 90% on the T1/T2 combined dataset.
- Medical experts validated the superiority of AM-UNet over other models.
Conclusions:
- AM-UNet is a highly accurate and efficient automated method for human brain claustrum segmentation.
- The model offers a robust solution for challenging CL segmentation tasks in neuroimaging.
- The source code and model are publicly available for research use.
