Development of Tomographic Reconstruction Method for MTV using Neural Radiance Fields

This work explores the use of a Neural Radiance Field–based tomographic reconstruction approach for 3D Molecular TaggingVelocimetry (MTV) measurements obtained with a plenoptic camera. Conventional tomographic reconstruction typically relies on voxel-based iterative algorithms, which can be sensitive to discretization, noise, and limited angular sampling. Here, the FluidNeRF framework is adapted to the Plenoptic 3.0 (Fourier Integral Microscope) imaging geometry by implementing an orthographic ray model and incorporating the elemental-image affine calibration directly into the reconstruction. The method is applied to an MTV dataset acquired in an axisymmetric stagnation jet experiment, consisting of 19 elemental images captured from a single camera. Three dimensional MTV intensity fields reconstructed using FluidNeRF are compared with results obtained using Richardson–Lucy Deconvolution. Both approaches recover the expected beamlet arrangement and its depth-dependent behavior. However, the FluidNeRF reconstructions exhibit reduced background noise and smoother beamlet representations, particularly in regions where the tagged signal is weak. These results indicate that neural implicit representations can be applied to plenoptic MTV data and provide a viable alternative to conventional iterative reconstruction methods for recovering volumetric MTV intensity fields.