Learning implicit light propagation from multi-flame projections for computed tomography of chemiluminescence

Area of Science:

• * Computational Physics
• * Optical Diagnostics
• * Chemiluminescence Imaging

Background:

• * Traditional computed tomography for chemiluminescence relies on weight matrix calculation and Algebraic Reconstruction Techniques (ARTs).
• * Existing methods face challenges with large storage requirements for weight matrices and potential inaccuracies from simplified imaging models.

Purpose of the Study:

• * To introduce a novel deep learning method, Multi-weight Encode Reconstruction Network (Multi-WERNet), for 3D chemiluminescence reconstruction.
• * To enable simultaneous reconstruction of 3D flame chemiluminescence by learning light propagation physics directly from multi-projections.

Main Methods:

• * Development of Multi-WERNet, a deep learning network designed to encode voxel weights and learn implicit light propagation physics.
• * Utilizing multi-projections from different flames to train the network for simultaneous 3D reconstruction.
• * Comparison of Multi-WERNet performance against established Algebraic Reconstruction Techniques (ARTs).

Main Results:

• * Multi-WERNet achieves reconstruction results comparable to ART, notably without the radial streaks often present in ART-based methods.
• * Reconstructions using 5 views with Multi-WERNet demonstrate superior performance over the ART method, especially when leveraging information from multiple flames.
• * The network exhibits robustness in reconstructing from imperfect projections, enhancing its practical applicability.

Conclusions:

• * Multi-WERNet effectively learns implicit light propagation physics, functioning as a transferable voxel weight encoder.
• * The proposed method offers a more robust and practical alternative to traditional ART for 3D chemiluminescence computed tomography.
• * Multi-WERNet shows significant potential for advancing flame diagnostics and other imaging applications requiring accurate 3D reconstruction.