Development of machine vision algorithms for radiomics and advanced particle therapy

The medical treatment and management of pancreatic cancer represents one of the most significant challenges in modern oncology. Pancreatic ductal ade- nocarcinoma is characterized by a high mortality rate, primarily due to the asymptomatic nature of its early stages and the rapid progression of metas- tases. Consequently, precise localization and delineation of the tumor volume are critical prerequisites for diagnosis, staging, and the planning of advanced therapies. Parallel to diagnostic challenges, therapeutic innovation in tumor treatment in generel is of main importance. Boron Neutron Capture Therapy (BNCT) has emerged as a promising technique for targeting cancers that are difficult to treat with classical methods. This binary therapy relies on the selective accumulation of 10B carriers in tumor cells, followed by irradiation with thermal neutrons. To ensure the safety and efficacy of BNCT, it is essential to monitor the in-vivo distribution of boron, which is achieved through Single Photon Emission Com- puted Tomography (SPECT). This modality detects the gamma rays emitted during the neutron capture reaction. This work proposes a synergistic approach to these challenges by developing Deep Learning techniques. In particular, it investigates the use of Convolutional Neural Networks, such as U-Net and Attention architectures, to automate two critical tasks: the semantic segmentation of pancreatic tumors from CT scans and the reconstruction of SPECT images for BNCT dosimetry. By integrating machine vision into the clinical workflow, this approach aims to improve the accuracy of tumor targeting and the quality of metabolic imaging. Medical imaging modalities, including CT and SPECT, are fundamentally based on the mathematical solution of the inverse problem. While the direct problem consists of predicting measurements from a known object, the inverse problem seeks to reconstruct the internal structure of an object from a set of external projections. In tomographic imaging, this relationship is mathemat- ically described by the Radon Transform, which maps a function defined in two-dimensional space to a collection of line integrals. Analytic reconstruction methods, such as Filtered Back Projection, use the Fourier Slice Theorem to invert this transform. However, the inverse problem is inherently ill-posed, as small perturbations in the measured data, often caused by noise, can result in significant artifacts in the reconstructed images. Although analytic methods are computationally efficient, they are particularly sensitive to the low signal- to-noise ratios that characterize SPECT acquisitions. To mitigate these limitations, Iterative Reconstruction Techniques, like for example Maximum Likelihood Expectation Maximization and Ordered Sub- sets Expectation Maximization, explicitly model the physical acquisition pro- cess. These algorithms operate by iteratively estimating the image, forward- projecting it through a system model, comparing the simulated projections with the measured data, and updating the estimate to minimize reconstruction er- ror. While iterative methods generally produce higher-quality images, they are computationally intensive and sensitive to modeling accuracy. In order to overcome the limitations of both traditional reconstruction tech- niques and manual segmentation procedures, Deep Learning methods are em- ployed. The principal architecture considered is the U-Net, a Convolutional Neural Network specifically designed for biomedical image segmentation. The U-Net follows a symmetric encoder–decoder structure, in which the contracting path progressively reduces spatial resolution while extracting hierarchical fea- tures that capture image context. The expanding path subsequently restores spatial resolution through up-sampling and convolutional operations. A defin- ing characteristic of this architecture is the presence of skip connections that link corresponding layers of the encoder and decoder, allowing high-resolution feature maps to be transferred directly across the network. This design pre- serves fine spatial details that are otherwise lost during pooling operations and enables precise boundary localization. Beyond convolutional architectures, attention-based models and Transformer networks are also explored. These approaches dynamically weight the relative importance of different input elements, providing a form of global contextual modeling that complements the local feature extraction capabilities of convolu- tional networks. The segmentation of the pancreas presents a particularly challenging prob- lem due to the organ’s high anatomical variability and the low contrast be- tween pancreatic tissue and neighboring organs such as the liver and stomach. Additionally, pancreatic tumors typically occupy only a very small fraction of the total abdominal volume, leading to a pronounced class imbalance that de- grades segmentation performance. To address these issues, first is explored a one-step pipeline, and then a two-step segmentation pipeline is studied. In one- step pipeline an high-resolution segmentation network, such as a 3D U-Net or a Super-Resolution Segmentation Network (SRSnet), is applied. In two-step pipeline, a coarse localization procedure is applied to the full abdominal CT volume using a 3D convolutional or regression-based model, like YOLO v11 model. The objective of this stage is to identify the approximate location of the pancreas by predicting a two-dimensional bounding box, rather than producing a detailed segmentation. In the second stage, the CT volume is cropped to the region of interest defined by the localization step, and a high-resolution segmen- tation network, such as a 2D U-Net is applied. This strategy allows the model to focus its computational resources on discriminating pancreatic parenchyma from tumor tissue while minimizing the influence of irrelevant background anatomy. The segmentation models are trained using the Medical Segmentation De- cathlon dataset and validated on the independent CPTAC-PDA dataset to as- sess their generalizability. Preprocessing steps include resampling all volumes to an isotropic spatial resolution and applying intensity windowing to enhance soft-tissue contrast. For the BNCT application, the goal is to reconstruct the spatial distribu- tion of 10B concentration from the 478 keV prompt gamma rays emitted during neutron capture events. The experimental setup, referred to as SPECT-BNCT tomography, consists of a Lanthanum Bromide scintillator coupled to Silicon Photomultipliers and equipped with a pinhole collimator. Due to the limited availability of experimental training data for this specific configuration, a syn- thetic dataset is generated using Geant4 and FLUKA Monte Carlo simulations. These simulations model the complete acquisition geometry, including the irra- diation room, shielding structures, and detector response, in order to produce realistic projection data from known phantom distributions. Two reconstruction strategies are investigated. The first employs an itera- tive approach based on the Ordered Subsets Expectation Maximization algo- rithm, implemented within the PyTomography framework. This method ex- plicitly incorporates the system matrix, including effects such as the collimator point spread function and resolution recovery. The second strategy relies on Deep Learning, using autoencoders and U-Net-based architectures trained on the simulated dataset. These models learn a direct mapping between raw back- projected images and the corresponding ground-truth phantom distributions, thereby avoiding the computationally intensive iterative loops required by tra- ditional methods and enabling fast, direct image reconstruction, indeed reducing most of the timing reconstruction. In addition to the core imaging studies, a supplementary investigation is conducted to characterize the Gamma Irradiation Facility (GIF++) at CERN. Geant4 simulations are used to map the spatial distribution of dose rates within the irradiation bunker. The simulation results are validated through comparison with experimental measurements obtained using gamma probes and ionization chambers under different source attenuation configurations. Overall, this work demonstrates the effectiveness of Deep Learning approaches in addressing complex challenges in radiomics and particle therapy. In the con- text of segmentation, the two-step pipeline provides a robust solution to the small-organ problem by decoupling coarse localization from segmentation, en- abling accurate delineation of both the pancreas and the tumor despite anatom- ical variability. This level of automation has the potential to significantly reduce the workload of radiologists and improve the consistency of treatment planning. In the domain of SPECT reconstruction, Deep Learning models trained ex- clusively on Monte Carlo generated synthetic data are shown to effectively sup- press noise and recover spatial resolution in low-count metabolic images, offering a fast and reliable alternative to traditional iterative techniques. The validation of both FLUKA and Geant4 simulation frameworks further supports the use of these experiments as reliable tests for future detector development and method- ological research. Ultimately, the integration of machine vision algorithms into the medical physics workflow represents a concrete step toward more personal- ized, accurate, and efficient cancer treatments.