Improving shear-wave speed analysis in penile vibro-ultrasound elastography using self-supervised deep learning-based IQ denoising

Ultrasound vibro-elastography (UVE) estimates tissue stiffness from externally induced shear waves. In cylindrical structures such as the penis, reflections and multidirectional interference produce complex wavefields that reduce acquisition reliability. The purpose of this research is to develop a deep learning-based framework to improve shear-wave speed (SWS) analysis.

A deep learning (3D U-Net) IQ-denoising framework is proposed to suppress reflected, interfering, and non-propagating wavefield components prior to physics-based inversion. The framework processes measured IQ wavefields and reconstructs SWS using the local phase velocity imaging with wave-based filtering (LPVI-WBF) algorithm. A two-stage quality-control pipeline classifies acquisitions as Good, Borderline, or Bad using thresholds τ_keep = 0.45 and τ_rel = 0.33. The network is trained on Cohort A (10 patients, 342 acquisitions) via self-supervised Noise2Self and evaluated on an independent Cohort B (6 patients, 902 acquisitions) and a simulation cohort (n = 90).

On Cohort B, deep learning (DL) rescued 376/403 Borderline acquisitions (93.3%) and preserved 89.4% of native-Good, raising the Pass rate from 38.7% to 91.8% (+53.1%). Rescued and native-Good SWS distributions were statistically indistinguishable: independent comparison (n = 227 vs 376) yielded δ = ±0.099 m/s (TOST p < 0.0001), and 198 matched pairs confirmed δ = ±0.161 m/s - both well below the ± 0.30 m/s clinical margin. In simulation, Borderline-acquisition reliability rose from 0.239 to 0.503 (n = 35; Δ = +0.264, p < 0.0001), with paired SWS equivalence at a ± 0.50 m/s margin.

Self-supervised deep learning-based IQ denoising acts as a reliability reconstructor for processing SWS data, expanding the acquisition population usable for physics-based inversion while leaving the physics-based LPVI-WBF inversion itself unchanged.