Fully automated IVUS image segmentation with efficient deep-learning-assisted annotation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Disease Research

Background:

• Intravascular ultrasound (IVUS) image segmentation is crucial for diagnosing and managing coronary artery disease.
• Deep learning (DL) methods show promise but are hindered by limited annotated datasets.
• Reducing annotation effort is key to clinical adoption of DL for IVUS segmentation.

Purpose of the Study:

• To develop an efficient deep learning framework for automated IVUS image segmentation.
• To significantly reduce the annotation effort required for training segmentation models.
• To achieve clinically acceptable segmentation performance with minimal data.

Main Methods:

• A two-branch deep learning network integrating spatial and channel-wise probability attention modules.
• Active learning and model output interaction to guide expert annotation.
• Segmentation Quality Assessment (SQA) to identify valuable images for annotation.
• Iterative fine-tuning on newly annotated data.

Main Results:

• Achieved state-of-the-art (SOTA) segmentation performance on coronary IVUS data.
• Required no more than 10% of the training data compared to traditional methods.
• Demonstrated significant reduction in manual annotation effort.
• Validated on 38,771 frames from 266 subjects using 5-fold cross-validation.

Conclusions:

• The proposed framework efficiently automates IVUS image segmentation.
• Active learning and SQA effectively minimize annotation burden while maximizing model performance.
• This approach facilitates the clinical application of DL for coronary artery disease assessment.