Structure and Intensity Unbiased Translation for 2D Medical Image Segmentation

Area of Science:

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

Background:

• Deep segmentation models face challenges due to data distribution gaps.
• Retraining models for new data distributions is costly and time-consuming.
• Clinical device-embedded algorithms are often unretrainable, worsening data gap issues.

Purpose of the Study:

• To address structural disparities in data distribution gaps for deep segmentation.
• To propose a novel image-to-image translation method that reduces structural differences.
• To improve the performance of deep segmentation models on diverse datasets.

Main Methods:

• Developed a Structure-Unbiased Adversarial (SUA) network for image-to-image translation.
• The SUA network incorporates a spatial transformation block to reduce structural gaps.
• An intensity distribution rendering module adapts the deformed structure to the target intensity distribution.

Main Results:

• The SUA method successfully transfers both intensity and structural content across multiple datasets.
• Experimental results demonstrate superior performance compared to existing methods in closing data gaps.
• The proposed approach enhances segmentation accuracy by mitigating structural and intensity disparities.

Conclusions:

• The SUA network offers an effective solution for bridging data distribution gaps in deep segmentation.
• This method is particularly valuable for clinical applications where model retraining is infeasible.
• SUA advances generative translation by addressing structural differences, not just intensity variations.