Use of semi-synthetic data for catheter segmentation improvement

Area of Science:

• Medical imaging
• Machine learning
• Surgical robotics

Background:

• Data collection and labeling for medical device segmentation are labor-intensive.
• Lack of sufficient, high-quality data hinders machine learning model performance in minimally invasive surgery.
• Accurate segmentation of medical devices like catheters is critical for patient safety.

Purpose of the Study:

• To develop an algorithm for generating semi-synthetic medical images to augment real datasets.
• To improve the accuracy and generalization of deep neural networks for catheter segmentation.
• To reduce the time and effort required for data labeling in medical imaging.

Main Methods:

• Developed a semi-synthetic image generation algorithm using real images and continuum robot kinematics for catheter modeling.
• Generated diverse datasets of heart cavities with artificial catheters.
• Trained and compared deep neural networks (modified U-Net) on real-only versus real and semi-synthetic datasets.

Main Results:

• Semi-synthetic data significantly improved catheter segmentation accuracy.
• A modified U-Net trained on combined datasets achieved a Dice similarity coefficient of 92.6 ± 2.2%, compared to 86.5 ± 3.6% for real data only.
• The use of semi-synthetic data reduced accuracy spread and improved model generalization.

Conclusions:

• Semi-synthetic data generation is an effective method to enhance machine learning model performance in medical device segmentation.
• This approach addresses data scarcity and labeling challenges in surgical applications.
• The proposed method leads to more robust and accurate segmentation models, benefiting minimally invasive procedures.