Interactive segmentation of medical images using deep learning.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Deep learning for medical image segmentation requires extensive labeled data, incurring high time and labor costs.
• Current interactive segmentation methods often use early fusion, leading to sparse interaction information and suboptimal performance.
• Existing algorithms overlook boundary information, negatively impacting segmentation accuracy.

Purpose of the Study:

• To develop an efficient interactive segmentation method for medical images that reduces labeling costs.
• To improve the utilization of user interaction information in segmentation models.
• To enhance segmentation performance by incorporating boundary information.

Main Methods:

• Proposed an early and late fusion strategy to effectively integrate user interaction and image features.
• Introduced a decoupled head structure to extract and leverage image boundary information.
• Developed a boundary loss function to enforce boundary constraints during segmentation.

Main Results:

• The proposed method significantly improved segmentation performance compared to the baseline across three medical datasets (Chaos, VerSe, Uterine Myoma MRI).
• Achieved notable improvements in NoC@80 (number of clicks to reach 80% IoU), with a score of 1.69 on the Chaos dataset.
• Reduced annotation time by over 50%, requiring only 2-3 clicks per image compared to 25 minutes for manual annotation.

Conclusions:

• The combined early and late fusion strategy effectively utilizes sparse interaction data.
• The decoupled head and boundary loss enhance the model's focus on precise object boundaries.
• This interactive segmentation approach offers a cost-effective and efficient solution for medical image annotation.