Generating Synthetic Labeled Data From Existing Anatomical Models: An Example With Echocardiography Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning enhances medical image analysis but requires extensive labeled data.
• Data acquisition is hindered by time-consuming manual annotation and inter-observer variability.
• Existing methods struggle with the scarcity and quality of annotated medical imaging datasets.

Purpose of the Study:

• To develop an automated pipeline for generating synthetic medical images with paired labels for deep learning.
• To overcome the limitations of manual data labeling in medical image analysis.
• To validate the efficacy of GAN-generated images for training segmentation models.

Main Methods:

• Utilized generative adversarial networks (GANs), specifically CycleGAN, to synthesize realistic ultrasound images from anatomical models.
• Developed a pipeline to automatically derive annotations from anatomical models and transform them into synthetic images.
• Trained a convolutional neural network (CNN) for left ventricle and left atrium segmentation using only synthetic 2D echocardiography images.

Main Results:

• Synthetic images generated by the CycleGAN pipeline demonstrated high realism and utility for training deep learning models.
• CNNs trained on synthetic data achieved median Dice scores of 91, 90, 88, and 87 on four unseen real ultrasound datasets.
• Performance of models trained on synthetic data matched or surpassed inter-observer variability on real datasets.

Conclusions:

• The proposed pipeline effectively generates synthetic ultrasound images that can replace real data for training deep learning models.
• This automated data generation approach significantly reduces reliance on manual annotation and addresses data scarcity.
• The method is broadly applicable to various medical imaging segmentation and landmark detection tasks across different modalities.