Related Experiment Video
Updated: Aug 2, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.8K
Segmentation-guided domain adaptation and data harmonization of multi-device retinal optical coherence tomography
Shuo Chen1, Da Ma2, Sieun Lee3
1School of Engineering Science, Simon Fraser University, Burnaby, BC, Canada.
Computers in Biology and Medicine
|April 23, 2023
Summary
Domain adaptation using Cycle-Consistent Generative Adversarial Networks (CycleGAN) significantly improves Optical Coherence Tomography (OCT) image segmentation generalizability. This method enhances segmentation accuracy and image quality without requiring target domain labels.
Area of Science:
- Medical Imaging
- Computer Vision
- Machine Learning
Background:
- Optical Coherence Tomography (OCT) images from different devices exhibit varying intensity profiles, hindering the generalizability of automatic segmentation models.
- Models trained on one domain often perform poorly on data from other devices, necessitating robust domain adaptation techniques.
- Ground-truth labels are typically scarce in the target domain, posing a challenge for traditional adaptation methods.
Purpose of the Study:
- To develop a domain adaptation method for improving the generalizability of OCT image segmentation across different devices.
- To leverage Cycle-Consistent Generative Adversarial Networks (CycleGAN) for enhanced domain adaptation, particularly when only source domain labels are available.
- To evaluate the impact of the proposed method on both segmentation performance and adapted image quality.
Main Methods:
- A two-stage pipeline was implemented: first, training a segmentation model on the source domain, and second, adapting target domain images to the source domain.
- CycleGAN was employed to translate images from the target domain to the source domain, enabling segmentation using the source-trained model.
- Ablation studies were conducted with various loss functions to assess their impact on segmentation and image quality metrics.
Main Results:
- The proposed model achieved a significant 46.24% improvement in segmentation Dice score compared to non-adapted methods.
- The model reached 87.4% of the upper limit of segmentation performance, demonstrating high accuracy.
- Image quality metrics (FID, KID scores) confirmed that adapted images with better segmentation also exhibited superior quality.
Conclusions:
- Segmentation-driven domain adaptation is effective for retinal imaging processing, enhancing cross-device generalizability.
- The method reduces the need for manual labeling in new domains, lowering labor costs.
- The approach provides insights into improving segmentation quality in unlabeled image domains by incorporating anatomical information.
