Ultrasound elastic modulus reconstruction using a deep learning model trained with simulated data.

Area of Science:

• Medical Imaging
• Biophysics
• Machine Learning

Background:

• Ultrasound (US) elastography non-invasively measures tissue stiffness by solving an inverse problem from US images.
• Traditional inverse problem methods in elastography are computationally intensive or sensitive to noise.
• Limitations hinder the speed and accuracy of current US elastography techniques.

Purpose of the Study:

• To develop and validate a deep learning (DL) approach for solving the inverse problem in US elastography.
• To recover the spatial distribution of elastic modulus from US-measured displacement fields.
• To overcome the limitations of traditional iterative and direct inverse problem techniques.

Main Methods:

• A U-Net-based deep learning neural network was developed.
• The network was trained using simulated data from a forward finite element model.
• Model performance was evaluated using Mean Squared Error (MSE) and Mean Absolute Percentage Error (MAPE) metrics.
• The model was further validated with phantom experiments and clinical data.

Main Results:

• The DL model accurately reconstructed modulus distributions with low MSE and MAPE (e.g., mean MAPE of 0.32% for hard inclusions).
• Phantom studies showed predicted modulus ratios aligning with expected values, confirming accuracy.
• The model demonstrated robust generalization capabilities across diverse simulated, phantom, and clinical datasets.

Conclusions:

• Deep learning models trained on diverse simulated data generalize effectively to phantom and real-world clinical data.
• This data-driven approach offers a promising alternative to traditional methods in US elastography.
• The study validates the potential of DL for accurate and efficient non-invasive tissue property quantification.