3D-deep optical learning: a multimodal and multitask reconstruction framework for optical molecular tomography.

Area of Science:

• Biomedical Imaging
• Computational Imaging
• Deep Learning

Background:

• Optical molecular tomography (OMT) is an emerging imaging technique.
• Current deep learning reconstruction algorithms lack universality for diverse objects and probes.
• This limitation hinders OMT development and application.

Purpose of the Study:

• To present a universal deep learning framework for OMT reconstruction.
• To overcome the limitations of existing OMT reconstruction algorithms.
• To improve the applicability of OMT across various imaging scenarios.

Main Methods:

• Introduced a multimodal and multitask reconstruction framework: 3D deep optical learning (3DOL).
• Decomposed OMT reconstruction into optical field recovery and luminous source reconstruction tasks.
• Employed a recurrent convolutional neural network incorporating anatomical and boundary optical data, alongside 2D axial information for object recognition.

Main Results:

• 3DOL achieves stable, high-quality reconstruction with few parameters by recovering the optical field under geometric constraints and segmenting luminous sources using a learnable Laplace operator.
• Demonstrated compatibility with diverse objects through numerical simulations, physical phantoms, and in vivo experiments.
• Showcased generalization capabilities across a wide spectral range (620-900 nm NIR-I window).

Conclusions:

• 3DOL significantly enhances the universality of OMT reconstruction.
• The framework effectively integrates multimodal data for improved image quality and object recognition.
• 3DOL shows promise for broad applications in biomedical imaging due to its spectral adaptability and compatibility.