Related Experiment Video
Updated: Sep 4, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.9K
Multiscale unsupervised domain adaptation for automatic pancreas segmentation in CT volumes using adversarial
Yan Zhu1, Peijun Hu2, Xiang Li3,4
1Engineering Research Center of EMR and Intelligent Expert System, Ministry of Education, College of Biomedical Engineering and Instrument Science, Zhejiang University, Hangzhou, China.
Medical Physics
|July 14, 2022
Summary
This study introduces an unsupervised domain adaptation method for automatic pancreas segmentation, improving model performance on new datasets without manual annotations. The approach enhances segmentation accuracy for unseen medical images, aiding clinical practice.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer-Aided Diagnosis
Background:
- Automatic pancreas segmentation is crucial for diagnosing and treating pancreatic diseases.
- Current models struggle with domain shift due to variations in imaging data.
- Manual annotation of medical images is time-consuming and expensive.
Purpose of the Study:
- To develop a novel unsupervised domain adaptation method for pancreas segmentation.
- To address challenges of limited annotations and domain shift in medical imaging.
- To improve the generalization of pancreas segmentation models to unseen data.
Main Methods:
- A 3D semantic segmentation model with attention and residual modules was designed.
- A multiscale progressively weighted structure was introduced for varied field of views.
- Adversarial learning with a multiscale discriminator was employed to learn domain-specific and domain-ambiguous information from labeled and unlabeled data.
Main Results:
- The proposed domain adaptation method significantly improved pancreas segmentation performance on a private target dataset.
- The Dice Similarity Coefficient (DSC) increased from 58.79% to 72.73% when trained on the NIH-TCIA dataset.
- DSC improved from 62.34% to 71.17% when transferred from the Medical Segmentation Decathlon (MSD) dataset.
Conclusions:
- The method effectively utilizes feature correlations across data domains to train segmentation models on unlabeled data.
- The approach enhances model generalization, enabling meaningful segmentation of unseen data.
- This technique holds potential for applying models trained on public datasets to unannotated clinical CT images, assisting radiologists.
