Beyond Human Variability: Deep Learning for Intravascular Ultrasound Segmentation With Noisy Labels.

Area of Science:

• Medical Imaging
• Artificial Intelligence

Background:

• Intravascular ultrasound (IVUS) provides crucial cross-sectional views of coronary vessels.
• Deep learning excels at semantic segmentation for IVUS interpretation.
• Challenges include expertise required and time-consuming manual annotation.

Purpose of the Study:

• To evaluate the impact of noisy labels on deep learning-based IVUS segmentation.
• To develop a clinically informed filter for assessing label quality.
• To enhance model performance in semi-supervised learning settings.

Main Methods:

• Analysis of noisy label effects on deep learning segmentation models.
• Development of a filter using Hausdorff distance for label quality assessment.
• Application of the filter in a semi-supervised learning framework.

Main Results:

• Segmentation performance significantly degrades with consistent noisy labels or poor boundary alignment.
• Limited performance degradation (1.92% Dice coefficient) observed when 50% of boundaries were correct, even with 20-pixel errors.
• Increased training data size mitigated adverse effects of noisy labels.
• The proposed filter improved segmentation performance in self-training.

Conclusions:

• Noisy labels can substantially impact deep learning segmentation accuracy in IVUS.
• A clinically informed filter can improve model robustness by addressing label quality.
• Findings offer practical guidance for creating high-quality training datasets for IVUS imaging deep learning models.