SSP-Net: A Siamese-Based Structure-Preserving Generative Adversarial Network for Unpaired Medical Image Enhancement

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Unpaired medical image enhancement is crucial in medical research.
• Deep learning methods struggle with low-quality datasets and lack of paired data.
• Existing techniques often fail to preserve structural details during enhancement.

Purpose of the Study:

• To propose a novel deep learning method for unpaired medical image enhancement.
• To address the limitations of low-quality training sets and the absence of paired data.
• To achieve structure-preserving enhancement for medical images.

Main Methods:

• A Siamese structure-based network (SSP-Net) with a dual input mechanism was developed.
• The method considers both target highlight (texture enhancement) and background balance.
• Generative adversarial network (GAN) principles were integrated for structure-preserving enhancement through adversarial learning.

Main Results:

• SSP-Net demonstrated effective performance in unpaired medical image enhancement.
• The proposed method successfully enhanced texture and balanced background contrast.
• Experimental results showed SSP-Net outperforming other state-of-the-art techniques.

Conclusions:

• SSP-Net offers a viable solution for unpaired medical image enhancement.
• The dual input mechanism and GAN integration contribute to superior performance.
• This approach advances the field by overcoming data limitations in medical image processing.