A hybrid deep learning framework for real-time speckle reduction and image enhancement on portable ultrasound systems

Speckle patterns in ultrasound images often obscure anatomical details, leading to diagnostic uncertainty. Recently, various deep learning-based techniques have been introduced to effectively suppress speckle; however, their high computational costs pose challenges for low-resource devices, such as portable ultrasound systems.

To address this issue, we introduce Edge Speckle Reduction and Image Enhancement (EdgeSRIE), a lightweight hybrid deep learning framework for real-time speckle reduction and image enhancement in portable ultrasound imaging.

The proposed framework consists of an unsupervised despeckling branch and a self-supervised deblurring branch trained separately in sequence with AdamW (learning rate 1 × 10^{- 4}) and an L2 loss. The integrated model was converted to an 8-bit deployment model using post-training quantization on a low-resource system-on-chip. Training used 1779 B-mode images (BUSI and HC18), validation used 77 raw acquisitions (PICMUS and CUBDL), and external evaluation used 94 EdgeFlow UH-10 bladder images. EdgeSRIE was compared with OSRAD, OBNLM, DIAE, DUNet, BRUNet, and USNet using contrast-to-noise ratio (CNR), speckle signal-to-noise ratio (SSNR), average gradient magnitude (AGM), and structural similarity index measure (SSIM); paired inference used two-sided Wilcoxon signed-rank tests with Holm correction and rank-biserial and Hedges' g effect sizes.

Across the four representative cases, post-training-quantized EdgeSRIE achieved the largest mean improvement in CNR (64.9 ± 21.0%) and the smallest mean decrease in AGM (-16.6 ± 33.2%). SSNR (109.8 ± 45.8%) was in the OBNLM-led range, whereas SSIM remained slightly lower than OSRAD and OBNLM on its native [0, 1] scale. The deployed model contained 17.67K parameters and reached 64.10 frames/s on the target hardware. In the case-matched analysis (n = 8), EdgeSRIE showed Holm-corrected significant advantages over all six baselines in CNR, over five of six in SSNR, and over four of six in AGM. For SSIM, EdgeSRIE was significantly higher than the four deep learning baselines but lower than OSRAD and OBNLM. Across the statistically significant comparisons, the magnitude-based effect sizes were large (Hedges' g ≥ 0.8).

These results support the feasibility of EdgeSRIE as a compact, deployment-oriented framework that balances speckle suppression, structural preservation, and real-time execution for portable ultrasound imaging.