Image Reconstruction Using Deep Learning for Near-Infrared Optical Tomography: Generalization Assessment.

Area of Science:

• Biomedical optics
• Medical imaging
• Neonatal care

Background:

• Hypoxic ischemia poses a critical threat to preterm infants, necessitating rapid diagnosis.
• Near-infrared optical tomography (NIROT) offers potential for monitoring brain oxygenation but faces challenges with high-dimensional data processing.
• Timely intervention is crucial due to a narrow therapeutic window.

Purpose of the Study:

• To evaluate a hybrid deep learning model for near-infrared optical tomography (NIROT) image reconstruction.
• To assess the model's performance and generalization capabilities using synthetic and phantom data.
• To determine the model's effectiveness in improving speed and accuracy for neonatal brain oxygenation monitoring.

Main Methods:

• Development of a hybrid convolutional neural network (CNN) for NIROT image reconstruction.
• Training the CNN model using synthetic data.
• Testing generalization using physical phantoms with diverse geometries and source-detector arrangements, distinct from training data.

Main Results:

• The hybrid CNN demonstrated substantial improvements in reconstruction speed.
• Enhanced localization accuracy was achieved even with varied measurement conditions.
• High image quality was maintained despite using unseen, non-spherical inclusion shapes and divergent surface topologies.

Conclusions:

• Hybrid deep learning models are effective for accelerating NIROT image reconstruction.
• The developed CNN shows robust generalization capabilities for diverse neonatal brain imaging scenarios.
• This approach holds promise for improving the early detection of hypoxic ischemia in preterm infants.