Related Experiment Video
Updated: Jun 13, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.7K
TOPOLOGY-PRESERVING DEEP SUPERVISION FOR 3D AXON CENTERLINE SEGMENTATION USING PARTIALLY ANNOTATED DATA.
Roshan Kenia1, Fin Amin1, Benjamin W Roop1
1MIT Lincoln Laboratory.
Summary
This study introduces a new method for brain connectivity analysis, improving axon centerline detection with less annotated data. Our approach achieves high accuracy using only 50% of annotations, significantly reducing manual effort.
Area of Science:
- Neuroscience
- Computer Vision
- Biomedical Imaging
Background:
- Accurate axon centerline detection is crucial for mapping brain connectivity.
- Manual annotation of 3D brain imagery is labor-intensive and expensive.
- Developing methods that utilize limited or incomplete annotations is essential.
Purpose of the Study:
- To develop an accurate axon centerline detection technique that requires less annotated data.
- To address the challenge of incomplete annotations in 3D brain imagery.
- To improve the efficiency of brain connectivity analysis.
Main Methods:
- A novel topology-preserving loss function was developed.
- A deep supervision paradigm was integrated into the training process.
- The method was trained and evaluated using 3D brain volumes with varying levels of expert annotations.
Main Results:
- The proposed training paradigm significantly outperforms existing methods.
- Comparable performance was achieved using only 50% of the required annotations.
- The baseline method required 75% of annotations to achieve similar results.
Conclusions:
- The developed method effectively reduces the need for extensive manual annotation in brain imaging.
- This approach accelerates the process of understanding brain connectivity and functionality.
- The topology-preserving loss and deep supervision enhance accuracy with limited data.
