Regression-based neural network for improving image reconstruction in diffuse optical tomography

Area of Science:

• Biomedical Imaging
• Medical Physics
• Artificial Intelligence

Background:

• Diffuse optical tomography (DOT) is a non-invasive imaging method using light to detect tissue anomalies.
• Traditional DOT image reconstruction involves computationally intensive inverse problem solving, limiting its clinical application.
• Developing efficient algorithms for DOT is crucial for advancing medical diagnostics.

Purpose of the Study:

• To develop a novel, efficient deep learning approach for solving the inverse problem in DOT for compressed breast geometry.
• To utilize a cascaded feed-forward neural network for reconstructing DOT images and visualizing breast tissues and anomalies.
• To evaluate the performance of the deep learning-based DOT (DL-DOT) system against analytical solutions.

Main Methods:

• A regression-based cascaded feed-forward deep learning neural network was designed to solve the DOT inverse problem.
• A Monte Carlo algorithm was used to simulate light propagation in compressed breast models, generating the dataset (forward process).
• Performance was assessed using Pearson correlation coefficient (R) and Mean Squared Error (MSE) metrics.

Main Results:

• The DL-DOT system demonstrated a ~30% improvement in R compared to the analytical solution, despite using a smaller dataset.
• The proposed neural network significantly outperformed the analytical solution in terms of MSE, indicating greater robustness.
• The developed deep learning model effectively visualizes breast tissues and anomalies.

Conclusions:

• The developed feed-forward deep learning network provides an efficient and robust solution for the DOT inverse problem in breast imaging.
• This AI-driven approach shows significant potential for improving the accuracy and speed of DOT for medical applications.
• The DL-DOT system is adaptable for clinical settings, offering a promising advancement in non-invasive breast anomaly detection.