Diffuse Optical Tomography (DOT) is an emergent medical imaging modality, which employs near-infrared (NIR) light as investigating signal to recover the distribution of optical coefficients for both diagnostic screening and therapeutic follow-up. DOT reconstruction is notoriously challenging: due to the combination of coherent, quasi-coherent and primarily non-coherent photons at NIR wavelengths and the limited number of reliable boundary measurements, recovering tissue optical properties is a severely ill-posed inverse problem. Regularization strategies are thus essential to obtain a stable and meaningful reconstruction. Traditionally, DOT reconstruction has relied on model-based algorithms derived directly from the underlying physics. These approaches typically employ ℓ2-norm (Tikhonov/ridge), ℓ1-norm (lasso), or Elastic Net regularization. Although effective in simplistic cases, these methods require fine tuning of regularization parameters and struggle to reconstruct complex spatial distributions of optical coefficients. The recent success of deep learning has shifted tomographic imaging from purely knowledge-driven to increasingly data-driven paradigms. Data-driven reconstruction is now widely researched in CT and MRI, yet its application to DOT remains in an early stage. In this work, we introduce a hybrid knowledge and data-driven framework for DOT reconstruction. To address the ill-posed nature of the problem, we exploit a generative model – specifically the decoder part of an autoencoder-type neural network – as a learnable prior to represent the solution space. Rather than relying on manually designed regularization terms, the generative model implicitly incorporates prior knowledge about the physical structure and correlations in optical distributions, enabling reconstructions that reflect the underlying statistical properties of the data. To enforce interpretable representations, a sequence of linear layers between the encoder and decoder is inserted to provide better interpolation capabilities by minimizing the rank of the covariance matrix of the latent space. Once trained, this autoencoder produces a low-rank latent space, yielding an effective low-dimensional representation of the data. Additionally, we embed the resolution of the forward problem, carried out via a Graph Neural Network (GNN), into the iterative optimization scheme that solves the inverse problem. At each iteration, the GNN provides updated physical insights, acting as a learned model correction mechanism that adjusts the wrong components of the latent space, while extracting physically meaningful features to support reconstruction updates. The proposed framework has been evaluated on multiple test cases, and the experimental results suggest that incorporating the learned prior leads to improved reconstruction accuracy, especially when compared to standard approaches. In conclusion, this work establishes a foundational framework for a novel class of hybrid approaches for addressing inverse problems in DOT. The results highlight the potential of combining deep generative modeling with learned physics-based solvers, paving the way towards more accurate and practical imaging tools for biomedical diagnostics.
La Tomografia Ottica Diffusa (DOT) è una modalità emergente di imaging biomedico, che sfrutta la luce nel vicino infrarosso (NIR) per ricostruire la distribuzione dei coefficienti ottici, utile sia per lo screening diagnostico che per il follow-up terapeutico. La ricostruzione DOT è notoriamente complessa: la combinazione di fotoni coerenti, quasi-coerenti e prevalentemente incoerenti alle lunghezze d’onda NIR, unita al numero limitato di misurazioni affidabili al bordo, rende la stima delle proprietà ottiche dei tessuti un problema inverso fortemente mal posto. Di conseguenza, l’adozione di strategie di regolarizzazione è essenziale per ottenere ricostruzioni stabili e significative. Tradizionalmente, la ricostruzione DOT si basa su algoritmi modellistici derivati direttamente dalla fisica sottostante. Questi metodi utilizzano tipicamente la norma ℓ2 (Tikhonov/ridge), la norma ℓ1 (lasso) o la regolarizzazione Elastic Net. Sebbene efficaci in scenari semplici, tali approcci richiedono un accurato fine-tuning dei parametri di regolarizzazione e incontrano difficoltà nella ricostruzione di distribuzioni spaziali complesse dei coefficienti ottici. Il recente successo del deep learning ha spostato l’imaging tomografico da un paradigma puramente knowledge-driven a uno sempre più data-driven. La ricostruzione basata sui dati è ormai ampiamente studiata in CT e MRI, ma la sua applicazione alla DOT è ancora in una fase iniziale. In questo lavoro, introduciamo un framework ibrido, guidato sia dalla fisica sia dai dati, per la ricostruzione DOT. Per affrontare la natura mal posta del problema, utilizziamo un modello generativo – in particolare la parte di decoder di una rete neurale di tipo autoencoder – come learnable prior per rappresentare lo spazio delle soluzioni. Invece di affidarsi a termini di regolarizzazione progettati manualmente, il modello generativo incorpora implicitamente informazioni a priori sulla struttura fisica e sulle correlazioni nelle distribuzioni ottiche, consentendo ricostruzioni che riflettono le proprietà statistiche sottostanti dei dati. Per garantire rappresentazioni interpretabili, inseriamo una sequenza di layer lineari tra l’encoder e il decoder, migliorando la capacità di interpolazione mediante la minimizzazione del rango della matrice di covarianza dello spazio latente. Una volta addestrato, l’autoencoder produce uno spazio latente a basso rango, fornendo una rappresentazione dei dati efficace e a bassa dimensione. Inoltre, integriamo la risoluzione del problema diretto, realizzata tramite una Graph Neural Network (GNN), nello schema iterativo di ottimizzazione che risolve il problema inverso. Ad ogni iterazione, la GNN aggiorna le informazioni fisiche, agendo come meccanismo di correzione del modello appreso e aggiustando le componenti errate dello spazio latente, mentre estrae caratteristiche fisicamente significative a supporto della ricostruzione. Il framework proposto è stato valutato su diversi casi test, e i risultati sperimentali mostrano che l’integrazione del prior migliora significativamente l’accuratezza della ricostruzione rispetto agli approcci tradizionali. In conclusione, questo lavoro definisce un framework fondamentale per una nuova classe di approcci ibridi per problemi inversi nella DOT. I risultati evidenziano il potenziale della combinazione di modelli generativi con risolutori fisici, aprendo la strada a strumenti di imaging più precisi e pratici per la diagnostica biomedica.
ADVANCING BIOMEDICAL IMAGING THROUGH GENERATIVE MODELING: A HYBRID KNOWLEDGE AND DATA-DRIVEN APPROACH TO INVERSE PROBLEMS IN DIFFUSE OPTICAL TOMOGRAPHY
SERIANNI, ALESSANDRA
2026
Abstract
Diffuse Optical Tomography (DOT) is an emergent medical imaging modality, which employs near-infrared (NIR) light as investigating signal to recover the distribution of optical coefficients for both diagnostic screening and therapeutic follow-up. DOT reconstruction is notoriously challenging: due to the combination of coherent, quasi-coherent and primarily non-coherent photons at NIR wavelengths and the limited number of reliable boundary measurements, recovering tissue optical properties is a severely ill-posed inverse problem. Regularization strategies are thus essential to obtain a stable and meaningful reconstruction. Traditionally, DOT reconstruction has relied on model-based algorithms derived directly from the underlying physics. These approaches typically employ ℓ2-norm (Tikhonov/ridge), ℓ1-norm (lasso), or Elastic Net regularization. Although effective in simplistic cases, these methods require fine tuning of regularization parameters and struggle to reconstruct complex spatial distributions of optical coefficients. The recent success of deep learning has shifted tomographic imaging from purely knowledge-driven to increasingly data-driven paradigms. Data-driven reconstruction is now widely researched in CT and MRI, yet its application to DOT remains in an early stage. In this work, we introduce a hybrid knowledge and data-driven framework for DOT reconstruction. To address the ill-posed nature of the problem, we exploit a generative model – specifically the decoder part of an autoencoder-type neural network – as a learnable prior to represent the solution space. Rather than relying on manually designed regularization terms, the generative model implicitly incorporates prior knowledge about the physical structure and correlations in optical distributions, enabling reconstructions that reflect the underlying statistical properties of the data. To enforce interpretable representations, a sequence of linear layers between the encoder and decoder is inserted to provide better interpolation capabilities by minimizing the rank of the covariance matrix of the latent space. Once trained, this autoencoder produces a low-rank latent space, yielding an effective low-dimensional representation of the data. Additionally, we embed the resolution of the forward problem, carried out via a Graph Neural Network (GNN), into the iterative optimization scheme that solves the inverse problem. At each iteration, the GNN provides updated physical insights, acting as a learned model correction mechanism that adjusts the wrong components of the latent space, while extracting physically meaningful features to support reconstruction updates. The proposed framework has been evaluated on multiple test cases, and the experimental results suggest that incorporating the learned prior leads to improved reconstruction accuracy, especially when compared to standard approaches. In conclusion, this work establishes a foundational framework for a novel class of hybrid approaches for addressing inverse problems in DOT. The results highlight the potential of combining deep generative modeling with learned physics-based solvers, paving the way towards more accurate and practical imaging tools for biomedical diagnostics.| File | Dimensione | Formato | |
|---|---|---|---|
|
phd_unimi_R13846.pdf
accesso aperto
Licenza:
Creative Commons
Dimensione
3.08 MB
Formato
Adobe PDF
|
3.08 MB | Adobe PDF | Visualizza/Apri |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/359816
URN:NBN:IT:UNIMI-359816