A novel GAN-based three-axis mutually supervised super-resolution reconstruction method for rectal cancer MR image.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Rectal cancer magnetic resonance (MR) imaging is crucial for diagnosis and treatment planning.
• Low-resolution axial plane images, common due to cost and time constraints, hinder accurate interpretation.
• Existing super-resolution methods often fail to consider all anatomical planes, yielding suboptimal results for rectal cancer MRIs.

Purpose of the Study:

• To enhance axial resolution in rectal cancer MR imaging for improved diagnostic accuracy.
• To address the limitations of current super-resolution techniques in multi-planar MR image reconstruction.
• To develop a method that provides high-resolution images essential for both visual interpretation and quantitative analysis.

Main Methods:

• Proposed a Generative Adversarial Network (GAN)-based three-axis mutually supervised super-resolution (SR) method.
• Implemented 1D intra-slice SR for sagittal and coronal planes, and inter-slice SR via slice synthesis in the axial direction.
• Introduced Depth-GAN for axial slice synthesis, incorporating depth information and mutual learning across axes.

Main Results:

• Achieved a Peak Signal-to-Noise Ratio (PSNR) of 34.62 and a Structural Similarity Index (SSIM) of 96.34% on a test set.
• Demonstrated superior performance compared to existing super-resolution methods through extensive ablation studies.
• Validated the effectiveness of the proposed three-axis mutually supervised SR approach.

Conclusions:

• The developed GAN-based method significantly enhances axial resolution in rectal cancer MR images.
• The three-axis mutual supervision and Depth-GAN integration improve the accuracy of super-resolution reconstruction.
• This technique offers a promising solution for overcoming the challenges posed by low-resolution MR images in rectal cancer assessment.