Related Experiment Video
Updated: Jun 24, 2025
09:33
Three-dimensional Rendering and Analysis of Immunolabeled, Clarified Human Placental Villous Vascular Networks
Published on: March 29, 2018
9.7K
Patient-specific placental vessel segmentation with limited data.
Gary Sarwin1, Jonas Lussi2, Simone Gervasoni2
1Computer Vision Lab, ETH Zurich, 8092, Zurich, Switzerland. gary.sarwin@vision.ee.ethz.ch.
Journal of Robotic Surgery
|June 4, 2024
Summary
Patient-specific deep learning models overcome data disparities in medical imaging. Generative adversarial networks create artificial data, significantly improving vessel segmentation accuracy with limited real-world samples.
Area of Science:
- Medical Imaging
- Machine Learning
- Computer Vision
Background:
- Machine learning models struggle with medical data due to distribution shifts from inconsistent acquisition and patient variability.
- This data disparity hinders the clinical translation of trained models, limiting their practical application in healthcare.
Purpose of the Study:
- To propose and evaluate a patient-specific deep learning approach for medical image segmentation using generative adversarial networks (GANs).
- To address the challenge of limited annotated data in clinical settings by supplementing real data with GAN-generated synthetic samples.
Main Methods:
- Developed a training pipeline for patient-specific segmentation networks using limited annotated data.
- Employed generative adversarial networks (GANs) to generate artificial samples to augment the training dataset.
- Demonstrated the approach on endoscopic video data from fetoscopic laser coagulation for twin-to-twin transfusion syndrome.
Main Results:
- Achieved an intersection over union (IoU) score of 0.60 with only 20 annotated images using the proposed pipeline.
- A standard deep learning approach required 100 images to reach a comparable performance level.
- Training with 20 annotated images without the GAN-pipeline yielded an IoU of 0.30, highlighting a 100% performance increase with the proposed method.
Conclusions:
- Generative adversarial networks (GANs) can effectively generate synthetic data to significantly improve patient-specific model performance in vessel segmentation.
- The proposed patient-specific training pipeline offers a viable solution for enhancing automated vessel segmentation and facilitating clinical implementation.
