Abstract
PURPOSE: Ultrasound (US) elastography is a technique for non-invasive quantification of material properties, such as stiffness, from ultrasound images of deforming tissue. The material properties are calculated by solving the inverse problem on the measured displacement field from the ultrasound images. The limitations of traditional inverse problem techniques in US elastography are either slow and computationally intensive (iterative techniques) or sensitive to measurement noise and dependent on full displacement field data (direct techniques). Thus, we develop and validate a deep learning approach for solving the inverse problem in US elastography. This involves recovering the spatial modulus distribution of the elastic modulus from one component of the US-measured displacement field. APPROACH: We present a U-Net-based deep learning neural network to address the inverse problem in ultrasound elastography. This approach diverges from traditional methods by focusing on a data-driven model. The neural network is trained using data generated from a forward finite element model. This simulation incorporates variations in the displacement fields that correspond to the elastic modulus distribution, allowing the network to learn without the need for extensive real-world measurement data. The inverse problem of predicting the modulus spatial distribution from ultrasound-measured displacement fields is addressed using a trained neural network. The neural network is evaluated with mean squared error (MSE) and mean absolute percentage error (MAPE) metrics. To extend our model to practical purposes, we conduct phantom experiments and also apply our model to clinical data. RESULTS: Our simulated results indicate that our deep learning (DL) model effectively reconstructs modulus distributions, as evidenced by low MSE and MAPE evaluation metrics. We obtain a mean MAPE of 0.32% for a hard inclusion and 0.39% for a soft inclusion. Similarly, in our phantom studies, the predicted modulus ratio aligns with the expected range, affirming the model's accuracy. These findings, alongside evaluations using the modulus ratio and contrast-to-noise ratio, confirm our DL model's robust generalization capabilities across diverse datasets. CONCLUSIONS: The presented work demonstrated that provided the simulated data are sufficiently diverse and representative of a wide variability, the algorithm trained on simulated data would generalize well to both phantom, as well as real-world clinical data.