Unrolled-DOT: an interpretable deep network for diffuse optical tomography.

Area of Science:

• Biomedical imaging
• Optical tomography
• Medical physics

Background:

• Imaging through scattering media is crucial for applications like breast tumor detection and functional neuroimaging.
• Time-of-flight diffuse optical tomography (ToF-DOT) offers high-resolution imaging but relies on computationally intensive reconstruction algorithms.
• Traditional algorithms can suffer from long runtimes and lower reconstruction quality due to model mismatch or improper tuning.

Purpose of the Study:

• To develop a faster and more accurate inverse solver for ToF-DOT using a data-driven approach.
• To improve image reconstruction quality by addressing model mismatch in ToF-DOT.

Main Methods:

• Developed 'Unrolled-DOT', a data-driven unrolled network utilizing the learned iterative shrinkage thresholding algorithm.
• Incorporated a refinement U-Net and Visual Geometry Group (VGG) perceptual loss to enhance reconstruction quality.
• Trained and tested the model on simulated and real-world data, benchmarking against physics-based and learning-based solvers.

Main Results:

• Unrolled-DOT achieved state-of-the-art performance on real-world data.
• Demonstrated over a 10x reduction in both runtime and mean-squared error compared to traditional physics-based solvers.
• Outperformed existing learning-based algorithms in speed and reconstruction accuracy.

Conclusions:

• Introduced a novel learning-based inverse solver for ToF-DOT, 'Unrolled-DOT'.
• Achieved significant improvements in both speed and reconstruction quality.
• This advancement holds promise for noninvasive biomedical imaging applications.