Interactive Segmentation Using U-Net with Weight Map and Dynamic User Interactions.

Area of Science:

• Medical image analysis
• Computer vision
• Deep learning for medical imaging

Background:

• Interactive segmentation enhances medical image analysis by incorporating expert input.
• Traditional methods often require precise and numerous user interactions, limiting efficiency.

Purpose of the Study:

• To develop a novel interactive segmentation framework with dynamically sized user clicks.
• To introduce a weighted loss function based on user-clicked regions for improved deep learning models.

Main Methods:

• Implemented an interactive U-Net (IU-Net) model utilizing foreground and background user clicks.
• Developed a weighted loss function where clicked regions generate a weight map for deep neural networks.
• Trained and validated on the BCV dataset, tested on spleen and colon cancer CT images from the MSD dataset.

Main Results:

• The proposed weighted loss function significantly improved segmentation accuracy compared to a standard U-Net.
• Applying dynamic user click sizes increased overall accuracy by 5.60% and 10.39% with single user interactions.
• Demonstrated enhanced performance on challenging medical imaging datasets.

Conclusions:

• Dynamic user click sizing in interactive segmentation offers a more efficient and accurate approach.
• The novel weighted loss function effectively leverages user input for deep learning-based segmentation.
• This framework shows promise for improving clinical workflows in medical image analysis.