Learnable Priors Support Reconstruction in Diffuse Optical Tomography

Alessandra Serianni, Alessandro Benfenati, Paola CausinPhotonics 2025 · DOI

TL;DR

This paper introduces a novel hybrid approach for Diffuse Optical Tomography (DOT) reconstruction, combining a learned forward model with a data-driven prior generated by a deep autoencoder. This method significantly improves the accuracy of reconstructing optical properties in biological tissues, addressing the inherent ill-posedness of DOT.

What Problem Does This Paper Address?

Diffuse Optical Tomography (DOT) is a promising non-invasive imaging technique that uses near-infrared light to map the optical properties of biological tissues, offering insights into conditions like cancer and stroke. However, DOT is notoriously challenging because light scattering within tissues makes the reconstruction process severely ill-posed and ill-conditioned. Traditional methods rely on mathematical models and regularization techniques, which can be computationally expensive and struggle with complex tissue structures. This paper aims to overcome these limitations by leveraging the power of deep learning to create more effective regularization through learned "priors."

Methods

The core of this work is a hybrid framework that combines two key deep learning components:

  • Learned Forward Model (Graph Neural Network - GNN): Instead of relying solely on analytical solutions for light propagation, the authors employ a Graph Neural Network (GNN) to learn the forward model. This GNN takes the estimated optical properties and outputs the predicted light fluence at the detectors. A key advantage highlighted is that the learned GNN parameters are resolution-independent, allowing for consistent performance across different grid resolutions.
  • Learnable Prior (Autoencoder Decoder): To address the ill-posedness, the paper proposes learning a "prior" that represents plausible solutions for the optical properties. This prior is generated by the decoder part of a custom-trained autoencoder. The autoencoder is trained to map a low-dimensional latent code to estimated optical absorption coefficients. Crucially, the latent space is constrained to be low-rank, which helps to regularize the solution and improve the generative capabilities of the model.

The overall reconstruction process involves an optimization loop where a latent code is iteratively adjusted to minimize the discrepancy between the GNN's predicted light fluence and the actual measured data. The learned prior guides this optimization, effectively acting as a data-driven regularizer.

Datasets

  • COULE Dataset: The authors utilize a modified version of the COULE (Contrasted Overlapping Uniform Lines and Ellipses) dataset. This dataset, originally designed for CT applications, was downsized and adapted to simulate DOT scenarios with contrasted ellipses and circular regions representing areas of augmented optical absorption.
  • Synthetic Data Generation: For training and testing, 51,000 samples were generated. Each sample included a ground truth absorption coefficient map, and corresponding light fluence measurements were accurately simulated using the Diffusion Approximation model via a Finite Element method. 50,000 samples were used for training the autoencoder prior, and 1,000 for testing. The same sets were used for training and evaluating the GNN.

Key Insights and Ideas

  • Data-Driven Priors for Ill-Posed Problems: The paper demonstrates that deep generative models (specifically, autoencoder decoders) can effectively learn complex, data-driven priors that significantly improve reconstruction in severely ill-posed inverse problems like DOT.
  • Low-Rank Latent Space for Regularization: Constraining the latent space of the autoencoder to be low-rank is a key technique for enhancing interpolation capabilities and reducing the impact of noise, leading to better reconstructions.
  • Resolution-Independent GNNs: The use of GNNs for the forward model offers a significant advantage by learning parameters that are not tied to a specific grid resolution, simplifying deployment and improving generalization.
  • Hybrid Knowledge and Data-Driven Approach: By combining a physics-informed GNN (learned forward model) with a data-driven prior, the framework benefits from both the interpretability of physical models and the powerful representation learning of deep neural networks.

Key Results

  • Improved Reconstruction Accuracy: The proposed hybrid framework, incorporating the learned prior, shows improved reconstruction accuracy compared to methods without a prior or using traditional variational approaches (Bregman, Elastic Net regularization).
  • Quantitative Metrics: For reconstructions using 20 light sources, the Structural Similarity Index Measure (SSIM) improved to 0.375 and the Mean Absolute Error (MAE) decreased to 0.008, compared to SSIM of 0.305 and MAE of 0.023 with a single light source.
  • Generalization Capability: The GNN demonstrated good generalization capabilities, performing accurately on a finer grid (64x64) even when trained on a coarser grid (32x32).
  • Qualitative Improvements: Visualizations show that the reconstructions with the learned prior are sharper and capture more details of the absorption coefficient distribution, especially in areas with high-intensity absorption.

Limitations and Future Work

  • Overlapping Contrast Regions: The authors acknowledge that the method struggles to capture fine details at the edges of overlapping contrast regions, resulting in some blurring.
  • Comparison to X-ray CT: While the results are considered remarkable for DOT, they are noted as not comparable to standard X-ray tomography, which offers significantly better resolution.
  • Future Directions: The paper suggests extending this hybrid approach to explore more complex scenarios and potentially enable more accurate and practical imaging tools for biomedical diagnostics. The authors advocate for further exploration at the intersection of deep generative modeling and learned physics-based solvers.

Summary updated

Abstract

Diffuse Optical Tomography (DOT) is a non-invasive medical imaging technique that makes use of Near-Infrared (NIR) light to recover the spatial distribution of optical coefficients in biological tissues for diagnostic purposes. Due to the intense scattering of light within tissues, the reconstruction process inherent to DOT is severely ill-posed. In this paper, we propose to tackle the ill-conditioning by learning a prior over the solution space using an autoencoder-type neural network. Specifically, the decoder part of the autoencoder is used as a generative model. It maps a latent code to estimated physical parameters given in input to the forward model. The latent code is itself the result of an optimization loop which minimizes the discrepancy of the solution computed by the forward model with available observations. The structure and interpretability of the latent space are enhanced by minimizing the rank of its covariance matrix, thereby promoting more effective utilization of its information-carrying capacity. The deep learning-based prior significantly enhances reconstruction capabilities in this challenging domain, demonstrating the potential of integrating advanced neural network techniques into DOT.
BibTeX
@inproceedings{serianni2025learnable,
  author = {Alessandra Serianni and Alessandro Benfenati and Paola Causin},
  title = {{Learnable Priors Support Reconstruction in Diffuse Optical Tomography}},
  booktitle = {Photonics},
  year = {2025},
  doi = {10.3390/photonics12080746},
}
Figure