Semi-supervised mp-MRI data synthesis with StitchLayer and auxiliary distance maximization

Area of Science:

• Medical Image Analysis
• Deep Learning
• Radiomics

Background:

• Annotated multimodal medical data, crucial for deep learning, is scarce and expensive.
• Multi-parameter MRI (mp-MRI), including ADC and T2w images, is vital for prostate cancer (PCa) diagnosis.
• Existing methods struggle with generating high-quality, paired mp-MRI data.

Purpose of the Study:

• To develop a novel semi-supervised adversarial learning framework for synthesizing realistic mp-MRI data.
• To generate synthetic Apparent Diffusion Coefficient (ADC) and T2-weighted (T2w) images containing clinically significant (CS) prostate cancer (PCa) lesions.
• To improve the robustness and visual quality of synthetic medical images.

Main Methods:

• A sequential synthesis approach generating ADC maps first, then T2w images using a U-Net.
• Semi-supervised training incorporating both paired and unpaired data to learn image distributions and relationships.
• Decomposition of image generation into sub-image tasks using a StitchLayer for enhanced robustness.
• Auxiliary Jensen-Shannon divergence maximization to ensure synthetic images contain distinguishable CS PCa lesions.

Main Results:

• The method successfully synthesizes diverse mp-MRI images with visually high-quality, realistic CS PCa lesions.
• Generated images maintain the correct paired relationship between ADC and T2w modalities.
• Significant improvements in visual quality and quantitative metrics compared to state-of-the-art adversarial learning methods.
• The approach effectively addresses data scarcity in medical imaging AI.

Conclusions:

• The proposed semi-supervised adversarial framework is effective for synthesizing high-fidelity mp-MRI data.
• This method can alleviate the bottleneck of limited annotated data in medical image analysis.
• The generated synthetic data holds potential for improving deep learning models in prostate cancer detection.