Training generative adversarial networks for optical property mapping using synthetic image data.

Area of Science:

• Biomedical Optics
• Computational Imaging
• Machine Learning

Background:

• Spatial Frequency Domain Imaging (SFDI) is a valuable technique for non-invasively mapping tissue optical properties.
• Accurate optical property maps are crucial for disease diagnosis and monitoring.
• Training deep learning models, such as Generative Adversarial Networks (GANs), requires large, well-annotated datasets, which are often challenging to acquire experimentally.

Purpose of the Study:

• To develop and validate a GAN for predicting optical property maps (absorption and scattering) from SFDI data.
• To leverage synthetic data generation using Blender for robust GAN training.
• To assess the performance of a synthetically trained GAN against experimentally trained models and real experimental data.

Main Methods:

• Utilized Blender, an open-source 3D software, to generate synthetic SFDI datasets simulating various tissue models, including diseased tissues with tumors and complex geometries like cylindrical organs.
• Trained a GAN on these synthetic datasets to predict optical property maps.
• Performed bi-directional cross-validation between synthetically trained and experimentally trained GANs, including a hybrid approach with 90% synthetic and 10% experimental data.

Main Results:

• The synthetically trained GAN achieved highly accurate reconstructions of optical properties from single SFDI images, with mean normalized errors of 1.0-1.2% for absorption and 1.1-1.2% for scattering.
• These results significantly outperform GANs trained solely on experimental data (approx. 10% error).
• Cross-validation demonstrated visually accurate results with errors of 6.3-10.3% for absorption and 6.6-11.9% for scattering when using a hybrid training approach, indicating strong domain transfer.

Conclusions:

• Synthetic data generation using Blender provides a flexible and powerful method for training GANs for SFDI optical property mapping.
• Synthetically trained GANs offer superior accuracy and enable the simulation of complex imaging scenarios not feasible with conventional methods.
• This approach holds significant potential for developing robust, fast, and accurate tools for clinical SFDI systems.