Dr Lucia Florescu PhD CPhys FHEA

Compton scatter tomography (CST) is an imaging technique that utilizes scattered x-ray radiation for 3D reconstruction of the electron density (ρe), which is otherwise not directly possible with transmission computed tomography (CT), thus providing additional information for material differentiation. In this study, we considered a CST scanning mechanism where a cone-beam source and an angularly-selective flat detector rotate in tandem around the sample, and applied a deep-learning diffusion model to reconstruct the electron density of the sample, with the 3D reconstruction performed slice-by-slice. This scanning mechanism could be readily implemented with current technologies for image-guided radiation therapy. To facilitate deep learning studies, 13,010 pairs of ground-truth 2D slice images of electron density and the corresponding detector readings were generated through numerical experiments. A conditional diffusion model was trained on a data subset to predict the velocity signal over a cosine noise schedule. We employed a U-Net architecture for the diffusion model, augmented with ConvNeXt blocks, and applied DDIM sampling over 100 timesteps. Evaluation on a testing data subset demonstrated mean PSNR ≈37.1 dB, SSIM ≈0.899, and RMSE ≈0.0160, with similar performance (2 to 7% deviation for all metrics) when evaluated under Poisson-noised detector readings.

Quantitative photoacoustic imaging (PAI) aims to determine the spatially varying optical absorption coefficient of a sample using the measured photoacoustic (PA) signals. When imaging tissue, this can be used to investigate the absolute concentration of the various constituent chromophores, such as oxy- and deoxyhaemoglobin. Supervised deep learning approaches have achieved promising results when trained to predict the absorption coefficient using synthetic datasets. However, models trained using synthetic data struggle to generalise to real data. Furthermore, very limited experimental data is available for this task, causing models trained using these data to overfit. The purpose of this study is to address these challenges using transfer learning. For this, convolutional neural networks (U-Nets) were pre-trained on a diverse synthetic dataset, created using 3D optical and acoustic modelling, and then fine-tuned on a publicly available experimental phantom dataset. When compared to U-Nets that were randomly initialised and trained on just the experimental dataset, the fine-tuned U-Nets achieved a ∼17% lower root mean squared error (RMSE) when predicting the optical absorption coefficient of the inclusions in the experimental phantom test dataset. This study also shows that, so long as the image formation process is the same for both training and testing data, and the training images are diverse, then U-Nets trained on synthetic data created from non-anatomical images are able to generalise to synthetic data created from an anatomically realistic mouse model.

We introduce the hybrid broken-ray tomography (HBRT) for three-dimensional imaging of weakly scattering systems. The HBRT utilizes fluorescent contrast agents and combines the principles and advantages of the broken-ray tomography and the non-reciprocal broken-ray tomography introduced by us previously. The HBRT uses angularly resolved intensity measurements at the incident and fluorescence wavelengths to reconstruct the attenuation and scattering coefficients at the excitation wavelength anywhere within the sample, as well as the attenuation coefficient at the fluorescence wavelength and the contrast agent concentration in the regions of contrast agent accumulation. The principles of HBRT have been validated by Monte Carlo simulations.

The clinical translation of nanoparticle-based treatments remains limited due to the unpredictability of (nanoparticle) NP pharmacokinetics—how they distribute , accumulate, and clear from the body. Predicting these behaviours is challenging due to complex biological interactions and the difficulty of obtaining high-quality experimental datasets. Existing AI-driven approaches rely heavily on data-driven learning but fail to integrate crucial knowledge about NP properties and biodistribution mechanisms. We introduce a multi-view deep learning framework that enhances pharmacokinetic predictions by incorporating prior knowledge of key NP properties such as size and charge into a cross-attention mechanism, enabling context-aware feature selection and improving generalization despite small datasets. To further enhance prediction robustness, we employ an ensemble learning approach, combining deep learning with XGBoost (XGB) and Random Forest (RF), which significantly outperforms existing AI models. Our interpretability analysis reveals key physicochemical properties driving NP biodistribution, providing biologically meaningful insights into possible mechanisms governing NP behaviour in vivo rather than a black-box model. Furthermore, by bridging machine learning with physiologically based pharma-cokinetic (PBPK) modelling, this work lays the foundation for data-efficient AI-driven drug discovery and precision nanomedicine.

The clinical translation of nanoparticle-based treatments remains limited due to the unpredictability of (nanoparticle) NP pharmacokinetics$\unicode{x2014}$how they distribute, accumulate, and clear from the body. Predicting these behaviours is challenging due to complex biological interactions and the difficulty of obtaining high-quality experimental datasets. Existing AI-driven approaches rely heavily on data-driven learning but fail to integrate crucial knowledge about NP properties and biodistribution mechanisms. We introduce a multi-view deep learning framework that enhances pharmacokinetic predictions by incorporating prior knowledge of key NP properties such as size and charge into a cross-attention mechanism, enabling context-aware feature selection and improving generalization despite small datasets. To further enhance prediction robustness, we employ an ensemble learning approach, combining deep learning with XGBoost (XGB) and Random Forest (RF), which significantly outperforms existing AI models. Our interpretability analysis reveals key physicochemical properties driving NP biodistribution, providing biologically meaningful insights into possible mechanisms governing NP behaviour in vivo rather than a black-box model. Furthermore, by bridging machine learning with physiologically based pharmacokinetic (PBPK) modelling, this work lays the foundation for data-efficient AI-driven drug discovery and precision nanomedicine.

The adoption of silicon photomultiplier (SiPM) detectors over conventional photomultiplier tubes (PMTs) in Positron Emission Tomography (PET) has enhanced overall system performance. In this phantom study, small-lesion detectability was assessed for SiPM-based and PMT-based PET systems for various inhomogeneity sizes, acquisition times and activity contrasts between the inhomogeneity and background. Six spheres of internal diameters ranging between 4.0 mm and 13.0 mm were integrated into a NEMA/IEC PET Body Phantom and filled with fluorodeoxyglucose, with a sphere activity concentration of 29.2 MBq/L and five sphere-to-background activity concentration ratios between 4 and 20. Scans were performed with an SiPM-based system and a PMT-based PET system for each sphere-to-background activity concentration ratio for acquisition times between 1 and 10 min, and image reconstruction was performed with QClear for both systems. Reconstructed images were evaluated for lesion detectability by a lesion detectability index, contrast-to-noise ratio and lesion detectability Likert scales with validation by comparison with the Rose criterion. A model to estimate the acquisition time for each sphere to be detectable was derived and acquisition time was compared. The SiPM-based system demonstrated superior lesion detectability, identifying smaller and less active spheres with shorter acquisition times. For a sphere-to-background activity concentration ratio of 10 and a sphere internal diameter of 6.2 mm, the SiPM-based system achieved a contrast-to-noise ratio of 15.8 and a lesion detectability Likert score of 3, compared to 12.0 and 2, respectively, for the PMT-based system. The acquisition time of the SiPM-based system could be reduced by between 1.6% and 89%, depending on sphere size and sphere-to-background activity concentration ratio. The minimum CNR required for a sphere to achieve a detectability Likert score of 0.5 was 6.3, consistent with the Rose criterion. SiPM-based PET has enhanced lesion detectability, especially for smaller, less active regions and for shorter acquisition times. A five-point Likert scale is an effective measure of lesion detectability. Guidance is also provided for choosing the acquisition time as a function of lesion size and activity uptake, and for changes in image quality testing protocols.

Purpose: Investigate Cherenkov light emission in intensity modulated radiation therapy (IMRT) of laryngeal cancer; identify detection configurations for tumour probing using measurements of Cherenkov light at the patient surface.Methods: Numerical experiments were performed using Monte Carlo simulations and clinical Computed Tomography (CT) and radiotherapy treatment plan data for a two-beam IMRT treatment. CT data indicated the tissue types, and plan data was utilised to simulate radiation delivery. Spectrally dependent values were assigned to the optical parameters (absorption, scattering, refractive index, anisotropy) of each tissue type. Dose calibration (the number of Monte Carlo events corresponding to a 1cGy dose at the isocenter) was obtained by simulating the delivery of a 10 x 10 cm2 radiation field to water. The spatial and spectral characteristics of Cherenkov light within the tissue and at the patient surface were determined, as well as the origin within the tissue of light emerging in various regions on the surface.Results: Emitted Cherenkov light is localized in the tissue in regions of high-dose delivery. The spectrum of Cherenkov light at the patient surface is consistent with the tissue optical absorption spectrum and presents a peak in the near-infrared region. Cherenkov light emitted within the gross tumour volume (GTV) and immediately surrounding tissue emerges at the patient surface on a well-defined, beam-independent region, called here “GTV spot”. Near-infrared light emerging on the GTV spot has comparable intensity with light emerging on other areas on the surface.Conclusion: Measurements of near-infrared light on the patient surface can potentially enable probing the GTV and surrounding tissue for monitoring tumour changes during the treatment course. Restricting the light measurements to the reduced-area GTV spot (that can be determined a priori through simulations) could lead to easier implementation of a Cherenkov-light-based functional tomographic imaging technology with the radiotherapy system.

Positron emission tomography (PET) is a widely used imaging modality for the diagnosis and treatment of oncologic diseases. In this study, we evaluated the performance of digital PET/CT systems using subcentimeter microsphere inserts in a NEMA IEC Body Phantom. The digital system was compared with a non-digital PET scanner using the same image reconstruction method. Results revealed that the digital system maintained higher detectability for smaller spheres with an average of 1 Likert score higher for lesions under 7.9 mm, indicating its ability to detect smaller lesions more effectively than the non-digital system. Furthermore, we observed that the drop-off in contrast recovery occurs at smaller microspheres in the digital PET system compared with that for a non-digital PET scanner. This suggests that digital PET may require the use of smaller spheres in image quality testing to ensure accurate comparison of performance between digital systems. This implies that digital systems can more accurately and effectively distinguish subtle differences in image intensity and spatial distributions of intensity, leading to improved lesion visibility and detection, which is likely due to the superior imaging characteristics offered by underlying detection technology.

Increased accessibility and recent developments in positron emission tomography (PET) detector technology have enhanced PET's role in radiotherapy treatment planning. This study investigates the efficacy of silicon photomultiplier (SiPM) PET systems for improving volume delineation. The study used a modified NEMA IEC Body Phantom filled with fluorodeoxyglucose (F-18-FDG) imaged using both a digital SiPM-based PET scanner and a non-digital, photomultiplier tube PET scanner. Results show distinct differences in target volumes determined using the two systems, with the digital system consistently demonstrating larger delineated target volumes between 1.91 - 3.56 times larger than that of the non-digital system. Target volumes delineated by the digital system were more reflective of the true geometric volume of the spheres with a range between 440 mm(3) and 984 mm(3) versus 209 mm(3) and 419 mm(3) for the non-digital system compared to the geometric volume of the spheres which was 2156 mm(3). This was most pronounced for higher sphere-to-background activity concentration ratios and smaller structures. This study suggests that digital PET allows for better selection of appropriate cancer treatment and could offer benefits for targeted radiation therapy.

Apparatus and methods for constructing a tomographic image of a sample are disclosed. The apparatus comprises at least one source configured to emit electromagnetic radiation at a first wavelength, at least one angularly-selective detector configured to detect the electromagnetic radiation at a second wavelength after the electromagnetic radiation has interacted with the sample, and a controller configured to construct a tomographic image of the sample based on information gathered using the at least one detector. The controller obtains information indicative of an intensity of the electromagnetic radiation detected at a second position by the at least one detector while the source is in a first position. Then, the source and detector positions are interchanged, and the controller obtains information indicative of an intensity of the electromagnetic radiation detected at the first position by the at least one detector while the electromagnetic radiation is emitted from the second position.

CAR-T cell immunotherapy is a promising technique for cancer treatment. To better understand and improve its efficacy for solid tumours, methods for in-vivo imaging and quantifying the CAR-T cell distribution are necessary. One approach involves inserting a reporter gene into the CAR-T cells, causing them to express photochromic proteins that provide strong near-infrared (NIR) optical contrast. NIR photoacoustic (PA) imaging is then used to image these proteins, and implicitly the CAR-T cells. The laser pulse in PA imaging causes a systematic and repeatable variation in the contrast provided by the photochromic proteins between successive scans that is distinguishable from the constant background contrast. In this study, machine learning (ML) techniques are used to classify and predict the spatial concentration of the proteins by analysing time-series PA images. To address the need for large training datasets, we developed a novel 3D simulation framework, which generates labelled PA images of CAR-T cells expressing the reporter gene. The framework was used to procedurally generate, and simulate imaging of, 629 digital samples, each of these was scanned sequentially by 32 laser pulses, resulting in 20,128 images. Neural networks, specifically a Multi-Layer Perception (MLP) and U-Net, were applied for the pixel-wise binary classification and regression of the reporter protein. These exceeded the performance of a Random Forest (RF) algorithm which was previously applied in another study using a small (n=3) in-vivo dataset. The U-Net achieved a coefficient of determination (R-2) of 0.96 and a root mean squared error (RMSE) of 4.3 x 10(-9) M, which represents a significant improvement when compared with the R-2 of 0.72 and RMSE of 1.1 x 10(-8) M achieved by the RF. This study proposes a potential advancement in the accurate non-invasive image detection and quantification of CAR-T cells, with the goal of accelerating preclinical research in cancer immunotherapy for solid tumours.

Significance Cherenkov light emitted in the tissue during radiation therapy enables unprecedented approaches to tumor functional imaging for early treatment assessment. Cherenkov light-based tomographic imaging requires image reconstruction algorithms based on internal light sources that, in turn, require knowledge about the characteristics of the Cherenkov light within the patient. Aim We aim to investigate the spatial and spectral characteristics of Cherenkov light within the patient and at the patient's surface, and the origin within the tissue of light reaching the surface, to provide insight for the development of image reconstruction algorithms for Cherenkov light-based tomographic imaging. Approach Numerical experiments using clinical patient data and Monte Carlo simulations are performed for the radiation therapy of laryngeal cancer for intensity-modulated radiation therapy and volumetric-modulated arc radiation therapy. Results The emitted Cherenkov light is concentrated in regions of high delivered dose, with the spatial distribution within the patient and at the patient's surface depending on the treatment type and patient anatomy. The Cherenkov light at the patient's surface is dominant in the near-infrared spectral region. Light emitted within the tumor emerges at the patient's surface on a well-defined radiation beam-independent region. The distribution within the patient of the emitted light that emerges on reduced areas on the patient's surface containing this region is similar to that of the light that emerges across the entire patient's surface. Conclusions Detailed information about the spectral and spatial characteristics of Cherenkov light is provided. In addition, these results suggest that surface light measurements restricted to smaller areas containing the region where the light emitted in the tumor emerges (that can be determined through simulations prior to the treatment) could enable probing the tumor while being easier to integrate with the radiotherapy system and while the effect of measurement data incompleteness on image reconstruction may not be too strong.

Quantitative photoacoustic imaging aims to determine the spatial distribution of the tissue’s optical absorption coefficient from photoacoustic (PA) signals measured at its surface. We combine large scale optical and acoustic modelling to estimate the optical absorption coefficient from simulated PA signal measurements using a band-limited transducer array that provides limited angular coverage. We validated our approach using a digital mouse atlas, and a PA imaging forward model which is based on the MSOT in-Vision 256TM system (iThera GmbH, Munich). We were able to recover the absorption coefficient when it was assumed that the scattering coefficient was known exactly, and that the digital phantom was an extrusion out of the 2D imaging plane. We then investigated how the performance was affected when these two assumptions were relaxed, and when substantial negative pressure artifacts were present in the reconstructed images.

Broken ray transforms (BRTs) are typically considered to be reciprocal, meaning that the transform is independent of the direction in which a photon travels along a given broken ray. However, if the photon can change its energy (or be absorbed and re-radiated at a different frequency) at the vertex of the ray, then reciprocity is lost. In optics, non-reciprocal BRTs are applicable to imaging problems with fluorescent contrast agents. In the case of x-ray imaging, problems with single Compton scattering also give rise to non-reciprocal BRTs. In this paper, we focus on tomographic optical fluorescence imaging and show that, by reversing the path of a photon and using the non-reciprocity of the data function, we can reconstruct simultaneously and independently all optical properties of the medium (the intrinsic attenuation coefficients at the excitation and the fluorescence frequency and the concentration of the contrast agent). Our results are also applicable to inverting BRTs that arise due to single Compton scattering.

Numerical experiments were performed to analyse the effect of data loss at the edges of the sample on the accuracy of optical-CT reconstruction, in the context of applications to radiation dosimetry.

We address the interior problem of computed tomography that occurs when projection data is only available for a region in the interior of the sample. In this case, it is not possible to accurately reconstruct the attenuation function even in the interior domain. We consider an algorithm for correcting the interior tomography reconstruction which is based on prior knowledge in the interior domain. This correction algorithm is evaluated by performing numerical experiments with the Shepp-Logan phantom for various amounts of data loss, noise in the available projection data, various values of the attenuation function known a priori, and various positions within the sample where the prior information is available. Good performance of the algorithm based on prior knowledge at one point is demonstrated in the case of noiseless data. In the presence of noise in the projection data, improvements in the reconstructed attenuation function are obtained based on prior knowledge at a number of points in the interior domain. The robustness of the correction algorithm to errors in the values of the attenuation function used as prior knowledge was also investigated.

We present a tomographic imaging technique based on angularly-selective measurements of fluorescent light that enables for the first time simultaneous reconstruction of the attenuation coefficient at two energies and of the contrast-agent concentration.

Optical methods of biomedical tomographic imaging are of considerable interest due to their non-invasive nature and sensitivity to physiologically important markers. Similarly to other imaging modalities, optical methods can be enhanced by utilizing extrinsic contrast agents. Typically, these are fluorescent molecules, which can aggregate in regions of interest due to various mechanisms. In the current approaches to imaging, the intrinsic (related to the tissue) and extrinsic (related to the contrast agent) optical parameters are determined separately. This can result in errors, in particular, due to using simplified heuristic models for the spectral dependence of the optical parameters. Recently, we have developed the theory of non-reciprocal broken-ray tomography (NRBRT) for fluorescence imaging of weakly scattering systems. NRBRT enables simultaneous reconstruction of the fluorophore concentration as well as of the intrinsic optical attenuation coefficient at both the excitation and the emission wavelengths. Importantly, no assumption about the spectral dependence of the tissue optical properties is made in NRBRT. In this study, we perform numerical validation of NRBRT under realistic conditions using the Monte Carlo method to generate forward data. We demonstrate that NRBRT can be used for tomographic imaging of samples of up to four scattering lengths in size. The effects of physical characteristics of the detectors such as the area and the acceptance angle are also investigated.

We perform numerical experiments based on Monte Carlo simulations and clinical CT data to investigate Cherenkov light emission in molecular radiation therapy of hyperthyroidism, and demonstrate that Cherenkov light-based dosimetry could be feasible.

Numerical experiments based on Monte Carlo simulations and clinical CT data are performed to investigate the spatial and spectral characteristics of Cherenkov light emission and the relationship between Cherenkov light intensity and deposited dose in molecular radiotherapy of hyperthyroidism and papillary thyroid carcinoma. It is found that Cherenkov light is emitted mostly in the treatment volume, the spatial distribution of Cherenkov light at the surface of the patient presents high-value regions at locations that depend on the symmetry and location of the treatment volume, and the surface light in the near-infrared spectral region originates from the treatment site. The effect of inter-patient variability in the tissue optical parameters and radioisotope uptake on the linear relationship between the dose absorbed by the treatment volume and Cherenkov light intensity at the surface of the patient is investigated, and measurements of surface light intensity for which this effect is minimal are identified. The use of Cherenkov light measurements at the patient surface for molecular radiation therapy dosimetry is also addressed.