Robustness study of noisy annotation in deep learning based medical image segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Deep networks achieve high performance in medical image segmentation using meticulously annotated data.
• However, medical imaging data often contains noisy annotations, and their impact on deep learning segmentation remains understudied.

Purpose of the Study:

• To investigate the effect of noisy annotations on deep learning-based mandible segmentation from CT images.
• To assess the robustness of deep learning models to varying degrees of label noise.

Main Methods:

• Collected 202 head and neck CT images with initial rough mandible annotations.
• Corrected annotations to establish a reference standard.
• Trained deep networks with varying ratios of noisy labels and tested segmentation performance.
• Validated models on public datasets.

Main Results:

• Networks trained with noisy labels exhibited decreased segmentation performance compared to those trained with reference standards.
• Performance generally improved as the proportion of noisy labels decreased.
• No significant performance difference was observed when noisy cases comprised 20% or less of the training data.
• Cross-dataset validation confirmed competitive performance of models trained with noisy data.

Conclusions:

• Deep learning models demonstrate a degree of robustness to noisy annotations in mandible segmentation from CT scans.
• Labeling quality is critical for deep learning applications in medical imaging.
• Future research should focus on leveraging limited high-quality annotations to enhance deep learning performance with noisy data.