AutoSOUL: Deep Convolutional Neural Network for Autonomous Selection of Ultrasound Speckle Tracking Parameters.
Authors
Abstract
In medical diagnostics, ultrasound strain elastography offers a noninvasive method to assess tissue stiffness, aiding in distinguishing tumors, cysts, and benign lesions from healthy tissue. A strain elastography framework employs a speckle-tracking technique to estimate the displacement between precompression and postcompression radio frequency (RF) frames, which is spatially differentiated to calculate the strain field. Second-order ultrasound elastography (SOUL) is a recently published speckle-tracking algorithm that optimizes a regularized cost function comprising data fidelity and regularization terms. However, the success of SOUL to produce high-quality strain images depends on the manual selection of regularization parameters, making it user-dependent and limiting its efficiency and usability. To mitigate this manual parameter selection, herein, we propose AutoSOUL, a deep convolutional neural network (CNN)-based framework, to autonomously select the optimal regularization parameters of SOUL. A customdesigned CNN was trained on simulated datasets and fine-tuned on phantom datasets to classify a strain image as acceptable or unacceptable and predict the optimal strain image and parameter set corresponding to a given precompression and postcompression RF frame set. When tested on simulated, phantom, and in vivo breast datasets, the ranges of the accuracy, precision, recall, and F1-scores produced by the trained CNN model were 77%-99%, 68%-98%, 97%-100%, and 81%-99%, respectively. The RMSE, MAE, SNR, CNR, and SR values calculated from the optimal strain images suggested by AutoSOUL were within 1.5%-4%, 6%-9%, 0%-30%, 0%-21%, and 0%-6%, respectively, of values obtained from the corresponding manually tuned strain images. AutoSOUL reduced the execution time of the otherwise required manual tuning and selection process of SOUL by a factor of 84.4. These results indicate that AutoSOUL has the potential to enhance the usability of ultrasound elastography, facilitating more efficient clinical evaluation of tissue abnormalities.