Lateral Resolution Enhancement of Ultrasound Images via Auto-Encoder Network

Ultrasound imaging is widely used for medical diagnostics, but its resolution is inherently constrained by factors such as wavelength, focal length, scan line density, and frame rate. A fundamental trade-off exists between lateral and temporal resolution, where increasing scan line density enhances spatial detail at the expense of reduced frame rates. This study explores the potential of deep learning, specifically an AutoEncoder-based approach, to enhance lateral resolution without sacrificing temporal resolution. The performance of the AutoEncoder is evaluated against traditional interpolation methods, including nearest, linear, and spline interpolation, using structural similarity (SSIM), peak signal-to-noise ratio (PSNR), multi-scale SSIM (MS-SSIM), and feature similarity (FSIM) metrics. The results demonstrate that the AutoEncoder outperforms interpolation methods, achieving the highest SSIM and FSIM, indicating superior structural preservation and feature retention. Additionally , the RF signal analysis shows that while the AutoEncoder maintains the overall waveform structure, minor amplitude and phase deviations exist. These findings suggest that deep learning-based super-resolution can effectively enhance lateral resolution while minimizing traditional resolution trade-offs.