Deep Learning-aided Spatially-Weighted Ultrasound Attenuation Estimation

Abstract

Quantitative ultrasound attenuation imaging shows promise for clinical applications. Among attenuation coefficient slope (ACS) estimation techniques, the Regularized Spectral Log Difference (RSLD) method enhances the trade-off between spatial resolution and estimation precision through total variation regularization, but it is negatively affected by changes in the backscatter coefficient (BSC). A recent method, Spatially- Weighted Image Fidelity and regularization Terms (SWIFT), introduced a spatially-weighted RSLD approach to minimize tissue interface artifacts, but still faces challenges in accurately representing media geometry. Here, we introduce a deep learning (DL)-aided SWIFT approach that leverages spectral log ratio information to compute spatially varying weights, aiming to reduce estimation bias and refine ACS delineation. A U-Net-based model was trained with k-Wave simulations to segment ellipsoidal inclusions against a homogeneous background. Weights were computed from the model's output following edge detection and dilation operations. In physical phantom targets, the proposed method reduced root-mean-square error by 72% and 31%, improved contrast-to-noise ratio by 240% and 37%, and increased intersection over union by 0.41 and 0.08, compared to RSLD and SWIFT methods, respectively. In vivo analysis of thyroid nodule revealed enhanced border delineation. These results illustrate the promise of DL-assisted techniques to enhance the accuracy of attenuation coefficient estimation.