High-resolution ultrasound imaging is critical ““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““““ in clinical diagnosis, enabling early detection of abnor- malities and precise assessment of anatomical structures. While super-resolution techniques have been widely ex- plored in medical imaging, most existing approaches are restricted to fixed or integer scaling factors. Arbitrary- scale super-resolution, especially for ultrasound images, remains largely unaddressed. This study presents a novel pipeline integrating a lightweight model, EliteNet, with ar- chitectural and training modifications to support arbitrary and asymmetric scaling for ultrasound images. A resize layer is introduced at the head of the network to accept user-defined scaling factors, and a two-step training strat- egy is employed to enhance output quality. In the first step, the model is trained using a hybrid loss combining Structural Similarity Index (SSIM) and Frequency Domain Loss (FDL). In the second step, only the final layer is updated using SSIM and L1 loss, preserving learned fea- tures while eliminating artifacts. A dedicated dataset was collected and augmented using flips and reflective padding to ensure structural consistency. Low-resolution images were synthesized using both symmetric and asymmetric scale factors. Our approach yields visually superior results and demonstrates better generalization across arbitrary scales. Quantitatively, it achieves a PSNR of 22.8018 and SSIM of 0.5947, outperforming existing baselines such as ArbRCAN, SRDNet, RDUNet, and ABPN. Extensive abla- tion studies validate the effectiveness of the loss config- uration and training strategy. This work lays foundational groundwork for adaptive, high-quality ultrasound imaging and opens opportunities for real-time, resource-efficient diagnostic applications.
