Machine learning algorithms

Ultrasound tomography (UST) is one of the key imaging methods used in both medical diagnostics and industry. It enables non-invasive visualization of internal structures of objects based on the analysis of sound wave propagation. Recent scientific studies have presented advanced machine learning algorithms that significantly enhance the quality of image reconstruction in transmission and reflection tomography, setting a new standard for accuracy and efficiency in processing measurement data.

The research team analyzed the effectiveness of six modern machine learning techniques, encompassing both classical statistical approaches and deep neural models. The study included methods such as Elastic Net—a regression model with regularization, artificial neural networks (ANN), convolutional neural networks (CNN), Gated Recurrent Units (GRU), Long Short-Term Memory (LSTM) networks, and transfer learning techniques using the deep architecture RESNET-50. Each model was optimized for hyperparameter selection and tested on a set of 5000 tomography data samples, allowing for an objective assessment of their potential.

*Example comparison of reference measurements (left) and reconstructed measurements (right) – Elastic Net in Transmission UST*

*Example comparison of reference measurements (on the left) and reconstructed measurements (on the right) – Artificial Neural Network in Reflective UST.*

The quality of the reconstruction was evaluated based on four recognized metrics: Mean Absolute Error (MAE), Mean Squared Error (MSE), PSNR (Peak Signal-to-Noise Ratio), and Structural Similarity Index (SSIM). The results clearly indicate the superiority of the Elastic Net model, which achieved the lowest MAE (0.0044) and the highest PSNR (35.42). Convolutional Neural Networks (CNNs) also demonstrated high precision, achieving an MAE of 0.0059 and a PSNR of 32.02. Importantly, all analyzed methods achieved very high structural similarity (SSIM ~0.9999), confirming their usefulness in reproducing the internal structures of objects.

Transfer learning using RESNET-50 and classical ANN performed slightly weaker in comparison, but their results were still satisfactory, confirming the usefulness of these methods in situations with limited availability of training data. The conducted experiments showed that the proper selection and configuration of the model are crucial for achieving high-quality reconstruction.

Visual comparisons of reference measurements with reconstruction results, including those using the Elastic Net algorithm, demonstrate that it is possible to faithfully reproduce the actual structure recorded by the tomographic system. The high agreement between the reconstruction and the real image emphasizes the potential of the presented approach in applications requiring high precision.

The discussed research highlights the tremendous possibilities offered by integrating machine learning techniques with ultrasonic tomography. The results open new perspectives in medical imaging, quality control in industry, and any field where fast and accurate reconstruction of internal structures is essential.

The full version of the article is available at:
link to publication