Reducing tissue characterization uncertainty in adaptive proton therapy through image synthesis and spectral imaging
- Autores: David Viar Hernández
- Directores de la Tesis: Ángel Torrado Carvajal (dir. tes.), Juan Antonio Vera Sánchez (codir. tes.)
- Lectura: En la Universidad Rey Juan Carlos ( España ) en 2025
- Idioma: español
- Tribunal Calificador de la Tesis: Antonio Francisco González López (presid.), Mónica García Vázquez (secret.), Paul Deak (voc.)
- Programa de doctorado: Programa de Doctorado en Tecnologías de la Información y las Comunicaciones por la Universidad Rey Juan Carlos
- Materias:
- Enlaces
- Tesis en acceso abierto en: TESEO
- Resumen
Adaptive proton therapy (APT) demands highly accurate tissue characterization to ensure that proton dose is delivered as planned despite potential anatomical changes. Unlike conventional computed tomography (CT), cone beam CT (CBCT) Hounsfield units cannot be directly converted to proton stopping power ratios (SPR), which are essential for calculating how far protons travel in tissue. This limitation introduces substantial uncertainty in range estimation and dose delivery. Minor errors in SPR mapping can lead to proton range overshoot or undershoot, so reducing these uncertainties is vital to maintain the proton therapy¿s precision and to avoid enlarging treatment margins. Consequently, there is strong motivation to improve on-board imaging so that it yields more reliable tissue information for APT, minimizing discrepancies in SPR and, thus, enhancing the safety and efficacy of treatments.
To address these challenges, this doctoral Thesis explores advanced image synthesis and spectral imaging techniques to reduce tissue characterization uncertainty in APT. The central objectives are to generate high-quality synthetic CT images and dual-energy CT (DECT) images from standard CBCT scans, and to calibrate the DECT data for direct, accurate SPR mapping. By creating synthetic images that closely mimic diagnostic-quality CT and by leveraging the dual-energy information for material differentiation, the aim is to enable more precise estimation of proton stopping powers on the fly during treatment adaptation. Achieving these goals would allow clinicians to replan or adjust proton therapy using the daily CBCT with confidence that the computed dose distribution remains accurate despite potential patient anatomical changes The first step to pursue these aims was to generate synthetic CT images from standard CBCT scans. A 3D Vision Transformer model was trained on paired CBCT¿CT datasets to learn volumetric corrections for CBCT artifacts and HU biases. This approach enhanced anatomical detail and HU fidelity, producing synthetic CTs whose SPR maps deviated by less than 5% from planning CT¿derived values¿within clinical tolerance for APT workflows.
Next, to leverage spectral information without acquiring an actual DECT, a conditional 3D Denoising Diffusion Probabilistic Model (DDPM) was introduced to synthesize two monoenergetic CT images (i.e., 80 kVp and 140 kVp) from the daily CBCT input. The diffusion process iteratively refined noisy CBCT volumes into high- and low-energy synthetic scans. Tissue characterization was markedly improved, as the dual-energy information allows for material-specific contrast that a single CT cannot provide. In comparisons, the new diffusion-based approach outperformed earlier deep learning methods in both noise suppression and structural fidelity, The resulting DECT pairs exhibited noise and artifact levels comparable to real DECT, enabling clear separation of tissues¿particularly those with similar density but different composition.
Finally, a DECT-based calibration was implemented to map dual-energy images directly to proton SPR. By estimating relative electron density and effective atomic number from the DECT, the method produced voxel-wise SPR maps with errors below 1% in phantom studies. The consistency of these results also improved¿there was a notable reduction in the variability (standard deviation) of SPR errors compared to traditional approaches, indicating a more robust and reliable mapping. Such precision in SPR directly reduces the uncertainty margin that must be applied in proton range calculations. When integrated into dose recalculations for adaptive plans, these calibrated DECT-derived SPR maps led to dose distributions that closely matched those based on ground-truth data, thereby improving confidence in adaptive dose planning Together, these contributions produce a suite of tools that transform on-board CBCT from a positioning modality into a quantitative resource for APT. By synthesizing CT and DECT images and applying physics-informed calibration, the methods expand the diagnostic and planning information available to radiation oncologists and physicists. Clinicians can thus make better-informed decisions¿adjusting plans daily with confidence that the adapted dose will match the intended distribution. Future work will focus on prospective clinical validation, real-time integration into treatment consoles, and extension to other anatomical sites. This research paves the way toward truly adaptive, image-guided proton therapy where uncertainty is minimized and clinical decision-making is maximally informed.
