Addressing Inconsistent Labeling With Cross Image Matching for Scribble-Based Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Weakly-supervised learning, particularly scribble annotation, is increasingly used for medical image segmentation to lower costs.
• Scribble annotations present challenges due to incompleteness, subjectivity, and lack of standardization, impacting segmentation performance.
• Existing methods struggle with the inherent noise and variability in scribble-based training data.

Purpose of the Study:

• To develop a robust method for medical image segmentation using noisy scribble annotations.
• To mitigate the negative effects of inconsistent and incomplete annotation data on model training.
• To improve the accuracy and stability of segmentation models trained with weakly-supervised data.

Main Methods:

• A reference set of class-specific tokens and pixel-level features was constructed from diverse images.
• A pixel-level feature matching strategy was employed to guide the learning process using the reference set.
• Smoothing and regression techniques were integrated to align pixel-level features across images.

Main Results:

• The proposed strategy effectively guided the learning process, especially with inconsistent scribbles.
• The method demonstrated improved ability to learn consistent patterns despite labeling imperfections.
• Experiments on three datasets showed superior segmentation accuracy and stability compared to state-of-the-art approaches.

Conclusions:

• The developed reference set and matching strategy significantly enhance medical image segmentation accuracy and stability.
• This approach offers a viable solution for leveraging weakly-supervised learning with imperfect annotations.
• The findings suggest a pathway towards more efficient and reliable automated medical image analysis.