TL;DR
This paper introduces NeuDOT, a novel diffuse optical tomography (DOT) method that leverages neural fields (NF) to achieve high-resolution 3D reconstruction of optical absorbance in scattering media. By moving away from voxel-based representations to a continuous NF encoding, NeuDOT overcomes limitations of traditional DOT and deep learning approaches, demonstrating submillimeter resolution and accurate reconstruction of complex 3D objects.
What Problem Does This Paper Address?
Imaging deep within scattering media, like biological tissues or foggy environments, is a significant challenge. Diffuse Optical Tomography (DOT) aims to address this by reconstructing the optical properties (specifically absorption) of an object. However, conventional DOT methods discretize the imaging volume into voxels, which limits spatial resolution and increases computational cost. Existing deep learning (DL) approaches often require large, pre-generated datasets, which can be expensive and time-consuming to create. This paper tackles the problem of achieving high-resolution, computationally efficient, and data-efficient DOT reconstruction in scattering media.
Methods
The core innovation of NeuDOT lies in its use of Neural Fields (NF) for representing the optical absorbance distribution. Instead of discretizing the volume into voxels, NeuDOT uses a Multi-Layer Perceptron (MLP) network to learn an implicit function that maps any 3D coordinate to its corresponding absorption coefficient. This continuous representation allows for arbitrary resolution.
Key methodological components include:
- Neural Field Representation: An MLP network encodes the absorption coefficient $\mu_a(r)$ as a continuous function of spatial coordinates $r=(x, y, z)$. Positional encoding is used to inject high-frequency spatial information into the network, enabling it to capture fine details.
- Physics-Based Forward Model: A Finite Element Method (FEM) solver is integrated to simulate diffuse light propagation, accounting for optical heterogeneity. This model is crucial for bridging the gap between measurements and the NF representation.
- Adaptive Meshing: To balance computational cost and accuracy, an adaptive meshing technique is employed. This refines the mesh density in regions of high optical heterogeneity, focusing computational resources where they are most needed.
- Gibbs Sampling: This Markov Chain Monte Carlo (MCMC) method is used to efficiently sample new nodal points for updating the mesh during the adaptive meshing process.
- Unsupervised Learning: NeuDOT is trained in a self-supervised manner, minimizing a loss function that penalizes the discrepancy between measured and predicted data, thus avoiding the need for large, pre-generated training datasets.
Datasets
The authors evaluated NeuDOT using:
- Phantom Experiments: Three types of phantoms were used:
- Experiment 1: A slab phantom with 2D planar patterns (letter 'T', university logo, two stripes) to evaluate lateral resolution.
- Experiment 2: A slab phantom with three high-absorption rods of varying heights to assess axial reconstruction.
- Experiment 3: A cuboid phantom containing a 3D virtual blood vessel structure to test reconstruction of complex biological-like geometries.
- Simulation Studies: Digital phantoms were used to simulate various 2D and 3D object cases, including letters and stripes, to validate performance under controlled conditions.
No new benchmark datasets were introduced; rather, existing experimental setups and simulated phantoms were utilized.
Key Insights and Ideas
- Continuous Representation is Key: The shift from discrete voxel-based representations to continuous neural field representations is a fundamental enabler for achieving higher resolution and overcoming the limitations of traditional DOT.
- Bridging Physics and Deep Learning: The successful integration of a physics-based FEM forward model with a neural network allows NeuDOT to learn from data while respecting physical principles, leading to more accurate and robust reconstructions.
- Adaptive Meshing for Efficiency: The adaptive meshing strategy is crucial for making high-resolution reconstruction computationally feasible by intelligently allocating resources to regions of interest.
- Unsupervised Learning Solves Data Bottleneck: By being unsupervised, NeuDOT bypasses the need for extensive, costly data generation, making it more practical for real-world applications.
Key Results
- Submillimeter Lateral Resolution: NeuDOT achieved submillimeter lateral resolution, significantly outperforming conventional Levenberg-Marquardt (LM) and CNN-based methods.
- Accurate 3D Reconstruction: The method accurately reconstructed complex 3D objects, including a virtual blood vessel model, to a depth of 14 mm.
- Improved Boundary Recovery: NeuDOT demonstrated sharper boundaries and more homogeneous intensity distributions compared to LM and CNN, which often produced over-smoothed or blurred results.
- Computational Efficiency: While requiring more memory than CNN, NeuDOT with adaptive meshing was significantly more memory-efficient than LM and achieved comparable reconstruction times to LM, while being much faster than its non-adaptive counterpart.
- Superior Performance Metrics: Quantitative evaluations (MSE, DSC, SSIM, PSNR) consistently showed NeuDOT outperforming LM and CNN across various experiments.
Limitations and Future Work
- Computational Cost: While NeuDOT with adaptive meshing is computationally efficient, the current implementation still cannot fulfill real-time 3D visualization requirements.
- Forward Model Accuracy: The diffusion approximation used in the FEM solver can collapse in regions close to the light source or when $\mu_a \ll \mu_s'$, suggesting that more precise (but time-consuming) methods like Monte Carlo simulations could be explored.
- Generalization to Real Tissues: The current experiments were conducted on phantoms. Future work aims to apply NeuDOT to more realistic and complicated cases, such as in vivo imaging of animals and human tissues.
- Broader Applications: The authors envision NeuDOT being applied to a wider range of applications, including non-invasive breast cancer diagnostics, obstacle recognition in automated driving, and plant root growth monitoring.
- Speeding up Training: Further research into neural fields is expected to further speed up the training process of NeuDOT.
Summary updated