Nikolaos Dikaios

Medical image reconstruction began with an exact mathematical result. In 1917, Radon showed how to recover a function from its line integrals, providing the mathematical basis for computed tomography, positron emission tomography, and single-photon emission computed tomography. This review follows the development of reconstruction from analytical inversion to modern deep learning. We first present the Radon transform, its attenuated and three-dimensional generalizations, and the corresponding inversion formulae, including filtered backprojection. We then review iterative and optimization-based methods, including maximum-likelihood expectation-maximization, variational regularization, and compressed sensing. The discussion proceeds to deep learning, covering convolutional and transformer architectures, algorithm unrolling, physics-informed networks, neural operators, and generative diffusion priors. Our central argument is that analytical, iterative, and learned reconstruction are different formulations of the same inverse problem rather than competing paradigms. Many learned components derive from classical formulations, and many classical operators admit learnable counterparts. We conclude by discussing open challenges related to theoretical guarantees, trustworthiness, data efficiency, federated learning, and regulatory validation. The most reliable learned systems are likely to retain the mathematical and physical structure of established reconstruction methods.

The year 1958 marked the beginning of a new era, as Rosenblatt introduced the Perceptron, a computational model for supervised learning through error-driven weight adaptation inspired by biological neurons. The backpropagation algorithm, introduced by Rumelhart et al. in 1986, transformed multilayer networks by allowing for weight updates. In 1997, Maass proposed spiking neural networks to emulate the temporal and sequential learning of biological neurons, while Schultz et al. proposed reinforcement learning by integrating brain-like reward mechanisms for decision making. The 2000s witnessed significant advancements, with the development of convolutional neural networks and recurrent neural networks inspired by brain’s hierarchical and recurrent mechanisms, respectively. More recently, in 2020, Feinman and Lake presented hybrid neurosymbolic models to capture the brain’s ability to learn from sensory data and perform structured reasoning. This work explores the remarkable progression of breakthroughs toward brain-inspired learning, aiming to offer a new perspective in the dynamic interplay between neuroscience and computer science.

Objective: Establishing control and specification limits for Volumetric Modulated Arc Therapy (VMAT) pre-treatment quality assurance (PTQA) is essential for streamlining PTQA workflows and optimizing plan complexity. This study aimed to develop and implement new methods and tools across treatment sites of varying complexity using multiple global and local gamma index criteria. Approach: 350 VMAT plans comprising brain, prostate, pelvis and head and neck treatments were retrospectively compiled. For each site, control limits were obtained using Statistical Process Control (SPC) along with heuristic methods (scaled weighted variance (SWV), weighted standard deviation (WSD), skewness correction (SC)). Specification limits were derived employing a new formalism aligned with the heuristic approaches. Calculations were performed under various global and local gamma index criteria using custom-built software (freely available at https://github.com/AEvgeneia/SPC_GUI_Scientific_Tool.git). Main results: WSD and SC control and specification limits were comparable, while SWV deviated with increasing complexity and stricter gamma index criteria. Conventional criteria (e.g. global 3%/2 mm) lacked sensitivity to detect subtle errors. Global 2%/1 mm and 1%/2 mm, and local criteria stricter than 3%/1 mm, met sensitivity requirements for low-complexity plans while maintaining clear separation between control and specification limits to identify plans with suboptimal delivery accuracy. High-complexity plans showed that global criteria stricter than 3%/1 mm and all evaluated local criteria are optimal, provided specification limits for the most stringent criteria remain clinically acceptable. Significance: A nuanced framework is provided for determining control and specification limits for gamma index passing rates, as well as corresponding thresholds for the mean gamma index, allowing for site-specific detection of suboptimal treatment plans. The open-source software tool developed can operationalise the proposed methodology facilitating the clinical adoption of advanced statistical methods. Site-specific thresholds could serve as inputs for machine learning and deep learning algorithms aimed at automating error detection and PTQA classification for plan complexity management.

Autoimmune demyelinating central nervous system (CNS) disorders encompass a wide array of clinical entities ranging from the organ specific Multiple Sclerosis (MS) to systemic autoimmune diseases (SADs) such as Systemic Lupus Erythematosus (SLE) and Sjogrens syndrome (SS). Despite international research efforts, distinction of these entities at clinical, imaging and laboratory level re-mains challenging, with almost 20% of patients being misdiagnosed with MS, out of which more than 50% carry the misdiagnosis for at least 3 years, while 5% are misdiagnosed for 20 years or more. This work aims to identify early biomarkers that can discriminate MS from other clinical entities that are MS mimickers. For this reason, we have meticulously curated a unique biobank with serum, DNA, RNA, and cerebrospinal fluid (CSF) samples from 396 treatment-naive patients who presented with a demyelinating episode and for which we have recorded over 300 clinical, serological, and imaging parameters. These patients have undergone follow up at 6 months and 12 months from their first de-myelinating episodes and have been categorised as follows: MS, SAD with CNS involvement, and demyelination with autoimmune features (DAF). A hybrid model for semi-supervised tabular classification is proposed that integrates a graph attention network with dynamic graph learning via Random Forest proximity and kNN graphs with tabular prior data fitted network for direct classification. The model also performed feature selection, self-supervised contrastive learning, self-training and data augmentation.

Federated learning (FL) is advancing cancer research by enabling privacy-preserving collaborative training of machine learning (ML) models on diverse, multi-centre data. This systematic review synthesises current knowledge on state-of-the-art FL in oncology, focusing on breast, lung, and prostate cancer. Unlike previous surveys, we critically evaluate FL’s real-world implementation and impact, demonstrating its effectiveness in enhancing ML generalisability and performance in clinical settings. Our analysis reveals that FL outperformed centralised ML in 15 out of 25 studies, spanning diverse models and clinical applications, including multi-modal integration for precision medicine. Despite challenges identified in reproducibility and standardisation, FL demonstrates substantial potential for advancing cancer research. We propose future research focus on addressing these limitations and investigating advanced FL methods to fully harness data diversity and realise the transformative power of cutting-edge FL in cancer care.

We introduce the augmented total variation () regularization method for Radon‐type inverse problems. Our novel approach incorporates a dual variable into the regularization process, thereby extending and essentially augmenting traditional regularization techniques. The proposed method is robust, requiring only one algorithmic iteration to achieve accurate reconstructions. Numerical experiments on a modified Shepp–Logan phantom demonstrate that the augmented regularization consistently yields higher structural similarity index metric (SSIM) values and lower mean absolute difference (MAD) values compared to filtered backprojection and standard regularization. These findings indicate that our method not only reduces reconstruction errors but also preserves structural details.

Background and Purpose: Establishing control and specification limits for Volumetric Modulated Arc Therapy (VMAT) pre-treatment quality assurance (PTQA) is essential for streamlining the PTQA workflow and optimizing treatment plan complexity. This study aimed to define control and specification limits for the PTQA process across treatment sites of varying complexity using multiple global and local gamma index (γ) criteria. Materials and Methods: 350 VMAT plans comprising brain, prostate, pelvis and head and neck treatments were retrospectively selected. For each site, the control limits were obtained using the Statistical Process Control (SPC) along with heuristic methods (scaled weighted variance (SWV), weighted standard deviation (WSD), skewness correction (SC)). Specification limits were derived employing a novel formalism developed in accordance with the heuristic methods. Calculations were performed under various global and local γ criteria using custom software developed for this purpose. Results: WSD and SC control and specification limits were comparable, while SWV progressively deviated with increasing complexity and more stringent γ criteria. Global 2%/1 mm and 1%/2 mm, and local criteria stricter than 3%/1 mm, met sensitivity requirements for low-complexity plans while maintaining clear separation between control and specification limits to identify plans with suboptimal delivery accuracy. High-complexity plans showed that global criteria stricter than 3%/1 mm and all evaluated local criteria are optimal, provided specification limits for the most stringent criteria remain clinically acceptable. Conclusions: Control and specification limits were established to detect delivery inaccuracies due to excessive plan complexity. The methodology and software developed can help centers evaluate their PTQA thresholds.

The identification of genes that can initiate and progress carcinogenesis through mutations is a challenging issue due to the sparsity of mutations and the high mutational heterogeneity between tumors of the same cancer type. While current methods mostly rely on finding recurrently mutated genes, this work aims to recover infrequently mutated genes that can contribute to cancer, by exploiting clustering analysis. To that end, we constructed a Network Graph of 8303 patients and 198 genes from single-point-mutation data, retrieved from the The Cancer Genome Atlas (TCGA), to find patient-gene groups with the parallel use of two separate methodologies: (a) one based on Barber’s modularity index, and (b) one based on network dynamics. A systematic analysis was applied over 23 statistically meaningful groups of 2037 patients spanning 22 cancer types, derived from both methodologies, using the Fisher’s exact test followed by the Benjamini-Hochberg false discovery rate method. This procedure recovered 32 known significantly mutated genes and 4 putative driver genes that are not identified by standard driver-gene identification algorithms used in the TCGA consortium analysis. This study has also revealed 14 statistically significant patterns not reported in the current literature, resulting in a total of 18 hidden oncogenic relations. Notably, among the results, skin cutaneous melanoma is related to IL7R. However, current literature suggests that only IL7R overexpression is linked to skin cancers. To the best of our knowledge, this is the first reported pan-cancer intra-clustering analysis. Significance Driver genes can initiate and progress carcinogenesis through mutations. Here, we aim to recover infrequently mutated driver genes by constructing a Network Graph of 8303 patients and 198 genes from single-point mutation data to obtain patient-gene groups with the parallel use of two separate methodologies. Robust statistical analysis over the resulting modules recovered 18 hidden oncogenic relations: 4 putative driver genes and 14 statistically significant mutational patterns. To the best of our knowledge, this is the first reported pan-cancer intra-clustering analysis.