Correcting Pseudo Labels with Label Distribution for Unsupervised Domain Adaptive Vulnerable Plaque Detection

Area of Science:

• Medical image analysis
• Deep learning
• Computer vision

Background:

• Unsupervised domain adaptation (UDA) is crucial for medical image analysis to address performance drops in deep neural networks on new datasets.
• Pseudo-labeling is a common UDA technique, but it suffers from low pseudo-label precision and noise, limiting its effectiveness.

Purpose of the Study:

• To propose a novel UDA framework that enhances pseudo-label precision and mitigates noise in medical image analysis.
• To improve the performance of deep neural networks on unlabeled target datasets.

Main Methods:

• Developed a UDA framework based on label distribution learning to correct noisy pseudo-labels.
• Formulated the problem as noise label correction by converting categorical pseudo-labels to distributions.
• Iteratively updated network parameters and label distributions to refine pseudo-labels before retraining the model.

Main Results:

• The proposed framework demonstrated effectiveness in improving detection performance for vulnerable plaques using intravascular optical coherence tomography (IVOCT) datasets.
• Experimental results confirmed the framework's ability to enhance the detection accuracy on unlabeled target images.

Conclusions:

• The novel label distribution learning-based UDA framework successfully addresses the challenges of pseudo-label noise in medical image analysis.
• This approach offers a promising solution for improving the robustness and generalizability of deep learning models in medical imaging applications.