Learning COVID-19 Pneumonia Lesion Segmentation From Imperfect Annotations via Divergence-Aware Selective Training.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate segmentation of COVID-19 pneumonia lesions is crucial for patient diagnosis and treatment.
• Deep learning models require extensive, expert-annotated data, which is scarce and time-consuming to acquire.
• Training with imperfect or noisy annotations poses a significant challenge for current segmentation methods.

Purpose of the Study:

• To develop a robust deep learning framework for segmenting COVID-19 pneumonia lesions using a combination of clean and noisy annotations.
• To address the limitations of data scarcity and annotation burden in medical image analysis.
• To improve the performance of segmentation models when trained with imperfect labels.

Main Methods:

• Proposed a dual-branch network architecture to process accurate and noisy annotations separately.
• Introduced a Divergence-Aware Selective Training (DAST) strategy to manage varying label qualities.
• Implemented regularization through dual-branch consistency for severely noisy samples and refined slightly noisy samples.

Main Results:

• The proposed method significantly outperforms standard training processes when learning from imperfect labels for COVID-19 pneumonia lesion segmentation.
• The framework demonstrates superior performance compared to state-of-the-art noise-tolerant methods across different percentages of clean labels.
• Achieved higher segmentation accuracy despite the presence of inaccurate annotations.

Conclusions:

• The developed approach effectively trains deep learning models for medical image segmentation using imperfect annotations.
• DAST strategy successfully leverages noisy labels while mitigating their negative impact, enhancing model robustness.
• This method offers a practical solution for improving diagnostic tools in scenarios with limited expert annotation resources.