Super-resolution reconstruction of ultrasound image using a modified diffusion model

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Ultrasound imaging is crucial in medical diagnostics.
• Super-resolution (SR) aims to enhance the detail and clarity of low-resolution images.
• Existing SR methods face challenges in preserving fine details in ultrasound images.

Purpose of the Study:

• To develop and evaluate a modified diffusion model for ultrasound image super-resolution (DMUISR).
• To improve the suitability of diffusion models for SR tasks in ultrasound imaging.
• To enhance the quality and detail of reconstructed ultrasound images.

Main Methods:

• DMUISR incorporates a multi-layer self-attention (MLSA) mechanism and a wavelet-transform based low-resolution image (WTLR) encoder.
• The model processes interpolated and magnified low-resolution ultrasound images to generate high-resolution outputs.
• Evaluation was performed on 1,334 fetal head-circumference ultrasound images (HC18 dataset) at 2×, 4×, and 8× magnification.

Main Results:

• DMUISR outperformed mainstream SR methods (Bicubic, VDSR, DECUSR, DRCN, REDNet, SRGAN) across all tested scales.
• Achieved superior image quality with clear structures and rich textures in both hard and soft tissues.
• Demonstrated significant improvements in learned perceptual image patch similarity (over 50% decrease), peak signal-to-noise ratio, and structural similarity index measure.

Conclusions:

• The modified diffusion model, DMUISR, offers superior performance for ultrasound image super-resolution.
• DMUISR effectively suppresses over-smoothing and mode collapse issues common in SR reconstruction.
• The model shows potential for significantly improving medical ultrasound imaging quality, with optimized computational efficiency.