Machine learning model with physical constraints for diffuse optical tomography.

Area of Science:

• Biomedical Optics
• Medical Imaging
• Machine Learning

Background:

• Diffuse optical tomography (DOT) reconstruction is challenging due to ill-posedness and noise.
• Existing DOT methods are susceptible to model mismatches and complex boundary conditions.
• Ultrasound guidance has not been previously integrated with machine learning for DOT reconstruction.

Purpose of the Study:

• To introduce a machine learning model with physical constraints (ML-PC) for improved DOT reconstruction.
• To combine ultrasound-guided DOT with machine learning for enhanced imaging accuracy.
• To evaluate the performance of the ML-PC method against traditional techniques.

Main Methods:

• Development of an auto-encoder neural network for DOT reconstruction.
• Incorporation of physical constraints into the machine learning model.
• Integration of ultrasound guidance with the ML-PC framework.

Main Results:

• The ML-PC method demonstrated superior accuracy compared to existing models in both phantom and clinical studies.
• In phantom studies, ML-PC reduced mean percentage error for absorption coefficient reconstruction.
• Clinical studies showed improved contrast between malignant and benign breast lesions using ML-PC.

Conclusions:

• The proposed ML-PC method significantly enhances DOT reconstruction accuracy and reliability.
• Ultrasound-guided DOT combined with ML-PC offers a promising advancement for medical imaging.
• This approach has the potential to improve diagnostic capabilities in various clinical applications.