Proposed Solution For Rapid Improvement of Ultrasound Images Using Self-Supervised Deep Learning: A Methodological Approach

Plane-Wave Imaging (PWI) allows for ultrafast acquisition rates, but sacrifices spatial resolution and contrast compared to coherent plane-wave compounding (CPWC). This article presents a computational pipeline based on a self supervised U-Net architecture designed to elevate the quality of single plane-wave (1-PW) acquisitions to levels equivalent to 75-angle compounding. To overcome the scarcity of in vivo data, we implemented a geometric augmentation strategy on the public PICMUS benchmark, expanding a single phantom to 200 spatially paired samples. The model is optimized through a novel two-stage training scheme that employs selective regularization (Spatial Dropout) and balances the mean squared error (MSE) with the multi-scale structural similarity (MS-SSIM) in an asymmetric 10:1 ratio. Results demonstrate effective suppression of noise and grating lobes, outperforming recent literature while using a fraction of the training data. Quantitatively, the network achieves increases of up to +7.3 dB (42.2%) in PSNR and +0.206 (43.2%) in MS-SSIM compared to the 1-PW input. This approach demonstrates that prioritizing pixel fidelity alongside struc tural guidance prevents excessive smoothing in limited-data regimes, enabling high-resolution real-time ultrasound without hardware modifications.