Reliable Mutual Distillation for Medical Image Segmentation Under Imperfect Annotations

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Computer Vision

Background:

• Convolutional Neural Networks (CNNs) excel in medical image segmentation but require extensive, accurately annotated data.
• Collecting precise annotations is labor-intensive, leading to the use of coarse, imperfect labels.
• Label noise in these imperfect annotations significantly degrades CNN performance.

Purpose of the Study:

• To develop a novel collaborative learning framework to address label noise in medical image segmentation using CNNs.
• To enhance the robustness and accuracy of segmentation models trained with imperfect annotations.
• To reduce the burden of manual data labeling in medical imaging.

Main Methods:

• A collaborative learning framework where two CNN segmentation models mutually assist each other.
• Utilizing complementary knowledge by having one model refine training data for the other.
• Employing reliability-aware knowledge distillation with augmentation-based consistency constraints.
• Incorporating joint data and model augmentations to maximize the use of reliable information.

Main Results:

• The proposed method significantly outperforms existing approaches in handling noisy annotations across different noise levels.
• Demonstrated improvement of nearly 3% Dice Similarity Coefficient (DSC) on the LIDC-IDRI lung lesion segmentation dataset with 80% noise ratio.
• Validation across two benchmark datasets confirms the superiority of the collaborative learning strategy.

Conclusions:

• The collaborative learning framework effectively mitigates the negative impact of label noise in medical image segmentation.
• The approach offers a practical solution for training accurate segmentation models with less demanding annotation efforts.
• This method holds promise for advancing automated medical image analysis in resource-constrained settings.