Monte Carlo-based data generation for efficient deep learning reconstruction of macroscopic diffuse optical...

Area of Science:

• Biomedical Optics
• Computational Imaging
• Machine Learning

Background:

• Deep learning (DL) models require extensive, diverse datasets for accurate image formation from sensor data.
• Challenges in diffuse optical imaging include a lack of public datasets and varied instrumentation, hindering DL model development.
• Significant progress has been made in developing computationally efficient light propagation models.

Purpose of the Study:

• To demonstrate the use of Monte Carlo (MC) simulation platforms for generating large, representative datasets for training DL models.
• To address the data scarcity issue in diffuse optical imaging research.
• To support the development of DL models for various optical imaging applications.

Main Methods:

• Leveraging MC platforms like "Monte Carlo eXtreme" and "Mesh-based Monte Carlo" to create synthetic datasets.
• Developing data generator pipeline strategies for diverse applications such as fluorescence optical topography, tomography, and single-pixel diffuse optical tomography.
• Training DL models using MC-generated in silico datasets and validating with experimental data.

Main Results:

• MC-based in silico datasets effectively trained DL models.
• Validated DL models demonstrated accurate and promising image formation performance.
• The proposed methods showed potential across a range of instrumentation designs and sample properties.

Conclusions:

• MC-based data generation pipelines can efficiently create large, representative datasets for DL model training.
• This approach overcomes limitations associated with experimental data acquisition in optical imaging.
• The developed pipelines are expected to foster rapid, robust, and user-friendly image formation across diverse applications.