[Study on the inverse problem of diffuse optical tomography based on improved stacked auto-encoder].

Area of Science:

• Biomedical Optics
• Medical Imaging
• Computational Science

Background:

• The inverse problem in diffuse optical tomography (DOT) is inherently ill-posed, limiting imaging accuracy and clinical utility due to slow computation.
• Traditional methods like the Levenberg-Marquardt (LM) algorithm struggle with speed and precision for DOT's complex inverse problems.

Purpose of the Study:

• To develop and validate an improved stacked auto-encoder (SAE) method for solving the DOT inverse problem.
• To enhance imaging accuracy, reduce computational time, and improve noise immunity in DOT compared to existing techniques.

Main Methods:

• A stacked auto-encoder (SAE) neural network was employed to address the DOT inverse problem.
• The SAE architecture was optimized to a single-output configuration to decrease computational load.
• Performance was evaluated against traditional SAE and Levenberg-Marquardt (LM) iterative methods.

Main Results:

• The improved SAE method achieved a 46.21% lower Mean Square Error (MSE) than traditional iterative methods and a 61.53% lower MSE than the traditional SAE.
• Image Correlation Coefficient (ICC) improved by 4.03% over the LM method and 18.7% over the traditional SAE.
• The proposed method demonstrated robust noise immunity under 3% noise conditions and reduced computation time to just 1.67% of the LM method's time.

Conclusions:

• The enhanced SAE method offers superior image quality and noise resistance for DOT inverse problems.
• This approach significantly accelerates computation, making it more suitable for clinical DOT applications.
• The study validates the effectiveness of neural networks, specifically improved SAEs, in advancing DOT imaging.