Efficient Click-Based Interactive Segmentation for Medical Image With Improved Plain-ViT.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Interactive medical image segmentation aims for precise results with minimal human input, crucial for diagnostics and prognosis.
• Click-based interactions offer an intuitive user interface compared to other methods.
• Existing models require further enhancement for interpreting click-based user feedback effectively.

Purpose of the Study:

• To propose a novel interactive segmentation framework leveraging iterative weighted loss and an enhanced Plain-ViT backbone.
• To improve the interpretation of user clicks for more accurate segmentation.
• To enhance segmentation performance and reduce interaction time.

Main Methods:

• Developed a Residual Multi-Headed Self-Attention encoder with hierarchical inputs and residual connections for the Plain-ViT backbone.
• Implemented an iterative weighted loss function guided by user clicks.
• Evaluated the framework on a prostate T2-MRI dataset and three public organ datasets.

Main Results:

• The proposed framework demonstrated superior performance compared to state-of-the-art methods.
• Achieved an 88.11% Intersection over Union (IoU) score on the prostate dataset.
• The iterative loss function training strategy accelerated model convergence, with 7.03 clicks for 80% accuracy.

Conclusions:

• The novel framework significantly enhances interactive medical image segmentation accuracy and efficiency.
• The proposed architecture and iterative loss function offer a robust solution for clinical applications.
• This approach represents a substantial advancement in automated and user-guided medical image analysis.