Improved Limited-View Ultrasound Tomography via Machine Learning

Area of Science:

• Medical imaging
• Nondestructive testing
• Geophysical exploration

Background:

• Tomographic reconstruction is vital across various scientific and industrial fields.
• Limited view (LV) configurations in tomographic imaging significantly degrade image quality due to restricted measurement angles.
• Existing LV compensation algorithms are computationally intensive or require application-specific tuning.

Purpose of the Study:

• To develop and evaluate a machine learning (ML)-based approach for limited view (LV) tomographic image reconstruction.
• To address the limitations of current compensation algorithms in terms of computational cost and customization.

Main Methods:

• An autoencoder (AE) architecture was employed for the ML-based LV compensation.
• The AE model was trained on an artificially generated dataset of LV and full-view tomographic images.
• The ML approach was evaluated against positivity regularization using laser-scanned corrosion maps, comparing RMSE and MAE.

Main Results:

• The ML-based approach demonstrated improved Maximum Absolute Error (MAE) in 80% of test cases compared to conventional methods.
• While conventional methods showed better mean Root Mean Squared Error (RMSE), the ML approach excelled in visual reconstruction quality, particularly for irregular features.
• The ML method achieved a 41% improvement in both RMSE and MAE when compared to raw LV images.

Conclusions:

• The proposed ML-based autoencoder offers a promising and effective solution for limited view (LV) tomographic reconstruction.
• This approach provides superior visual image quality and error reduction for complex structures compared to traditional methods.
• The ML technique presents a viable alternative to computationally expensive or bespoke compensation algorithms.