Related Experiment Video
Updated: Aug 22, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.9K
Latent shape image learning via disentangled representation for cross-sequence image registration and segmentation.
Jiong Wu1, Qi Yang2, Shuang Zhou3
1School of Computer and Electrical Engineering, Hunan University of Arts and Science, Changde, 415000, Hunan, China. wujiong@huas.edu.cn.
International Journal of Computer Assisted Radiology and Surgery
|November 8, 2022
Summary
This study introduces Latent Shape Image Learning (LSIL), a novel method for cross-sequence magnetic resonance image (MRI) registration and segmentation. LSIL effectively addresses domain shifts, showing superior performance in medical image analysis tasks.
Area of Science:
- Medical Imaging
- Computer Vision
- Machine Learning
Background:
- Cross-sequence magnetic resonance image (MRI) registration and segmentation are crucial for medical image analysis.
- Domain shifts between different MRI sequences present significant challenges.
Purpose of the Study:
- To propose a novel method, Latent Shape Image Learning (LSIL), using disentangled representations for cross-sequence MRI registration and segmentation.
- To address the challenges posed by domain shifts in medical image analysis.
Main Methods:
- Images were decomposed into shared domain-invariant shape and domain-specific appearance spaces using unsupervised image-to-image translation.
- A Latent Shape Image Learning (LSIL) model was developed on disentangled shape representations to generate latent shape images.
- Experiments involved cross-sequence image registration and segmentation, evaluated using Dice Similarity Coefficient (DSC) and 95th percentile Hausdorff Distance (HD95).
Main Results:
- The proposed LSIL method demonstrated superiority over state-of-the-art approaches on two datasets comprising 50 MRIs.
- The method achieved mean DSCs of 0.711 for cross-sequence registration and 0.867 for cross-sequence segmentation.
Conclusions:
- The novel representation disentangling method (LSIL) effectively solves cross-sequence registration and segmentation problems.
- Experimental results validate the feasibility and generalization of LSIL, highlighting its potential for clinical applications with missing sequences.
- The source code for LSIL is publicly available.
