Modeling and Synthesis of Breast Cancer Optical Property Signatures With Generative Models

Area of Science:

• Biomedical Optics
• Medical Imaging
• Computational Pathology

Background:

• Optical property quantification is a key biomedical imaging technique for tissue characterization.
• Translating optical properties to clinical pathology is challenging due to variability and light scattering.
• Advanced algorithms are needed to model complex tissue pathologies.

Purpose of the Study:

• To develop a data-driven, nonlinear model for real-time breast cancer pathology assessment using optical properties.
• To establish deterministic relationships between optical measurements and pathophysiology in an unsupervised manner.
• To enable automated margin assessment in surgical resections.

Main Methods:

• Utilized spatial frequency domain imaging to derive tissue optical properties.
• Employed a series of deep neural network models to create latent embeddings.
• Developed self-explanatory models to translate absorption and scattering properties and synthesize new data.

Main Results:

• Achieved rapid optical property modeling with errors comparable to current semi-empirical models.
• Demonstrated the ability to synthesize new data and systematically understand dataset properties.
• Successfully tested the method on 70 resected breast tissue samples (137 regions of interest).

Conclusions:

• The developed deep learning model effectively relates optical data signatures to underlying tissue pathology.
• This approach paves the way for deep automated margin assessment algorithms using optical imaging.
• The method offers a systematic understanding of optical properties and enables mass sample synthesis.