Harmonizing Pathological and Normal Pixels for Pseudo-Healthy Synthesis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Generative Adversarial Networks (GANs) show promise for synthesizing pseudo-healthy medical images.
• Existing GAN discriminators struggle to accurately identify lesions, limiting the quality of synthetic images.
• Accurate synthesis of pathology-free images is crucial for algorithm development and clinical applications.

Purpose of the Study:

• To develop an improved method for synthesizing pathology-free medical images from pathological inputs.
• To enhance the accuracy of lesion identification and the visual quality of synthesized images.
• To improve medical image enhancement and address low contrast issues in segmentation.

Main Methods:

• A novel Generative Adversarial Network (GAN) architecture incorporating a segmentor as a discriminator was developed.
• The generated pseudo-healthy images were utilized for medical image enhancement.
• A new metric evaluating synthetic image health based on label noise attributes was proposed.

Main Results:

• The proposed segmentor-based GAN significantly outperforms state-of-the-art methods in generating pseudo-healthy images.
• The method achieves superior performance using only 30% of the training data compared to existing approaches.
• Effectiveness was demonstrated on BraTS (T1 and T2 modalities) and LiTS datasets.

Conclusions:

• The segmentor-based GAN offers a powerful approach for generating high-quality, pathology-free medical images.
• This method improves medical image enhancement and segmentation, particularly in low-contrast scenarios.
• The approach is efficient, requiring less training data while achieving state-of-the-art results.