Deploying and processing neural representations of signals

Neural representations (NR) have emerged in the last few years as a powerful tool to represent signals from several domains, such as images, 3D shapes, or audio. Indeed, deep neural networks have been shown capable of approximating continuous functions that describe a given signal with theoretical infinite resolution. This finding allows obtaining representations whose memory footprint is fixed and decoupled from the resolution at which the underlying signal can be sampled, something that is not possible with traditional discrete representations, e.g., grids of pixels for images or voxels for 3D shapes. During the last two years, many techniques have been proposed to improve the capability of NR to approximate high-frequency details and to make the optimization procedures required to obtain NR less demanding both in terms of time and data requirements, motivating many researchers to deploy NR as the main form of data representation for complex pipelines. Following this line of research, we first show that NR can approximate precisely Unsigned Distance Functions, providing an effective way to represent garments that feature open 3D surfaces and unknown topology. Then, we present a pipeline to obtain in a few minutes a compact Neural Twin® for a given object, by exploiting the recent advances in modeling neural radiance fields. Furthermore, we move a step in the direction of adopting NR as a standalone representation, by considering the possibility of performing downstream tasks by processing directly the NR weights. We first show that deep neural networks can be compressed into compact latent codes. Then, we show how this technique can be exploited to perform deep learning on implicit neural representations (INR) of 3D shapes, by only looking at the weights of the networks.

Numerical simulations of the spine for patient-specific models of severe scoliosis

In silico methods, such as musculoskeletal modelling, may aid the selection of the optimal surgical treatment for highly complex pathologies such as scoliosis. Many musculoskeletal models use a generic, simplified representation of the intervertebral joints, which are fundamental to the flexibility of the spine. Therefore, to model and simulate the spine, a suitable representation of the intervertebral joint is crucial. The aim of this PhD was to characterise specimen-specific models of the intervertebral joint for multi-body models from experimental datasets. First, the project investigated the characterisation of a specimen-specific lumped parameter model of the intervertebral joint from an experimental dataset of a four-vertebra lumbar spine segment. Specimen-specific stiffnesses were determined with an optimisation method. The sensitivity of the parameters to the joint pose was investigate. Results showed the stiffnesses and predicted motions were highly depended on both the joint pose. Following the first study, the method was reapplied to another dataset that included six complete lumbar spine segments under three different loading conditions. Specimen-specific uniform stiffnesses across joint levels and level-dependent stiffnesses were calculated by optimisation. Specimen-specific stiffness show high inter-specimen variability and were also specific to the loading condition. Level-dependent stiffnesses are necessary for accurate kinematic predictions and should be determined independently of one another. Secondly, a framework to create subject-specific musculoskeletal models of individuals with severe scoliosis was developed. This resulted in a robust codified pipeline for creating subject-specific, severely scoliotic spine models from CT data. In conclusion, this thesis showed that specimen-specific intervertebral joint stiffnesses were highly sensitive to joint pose definition and the importance of level-dependent optimisation. Further, an open-source codified pipeline to create patient-specific scoliotic spine models from CT data was released. These studies and this pipeline can facilitate the specimen-specific characterisation of the scoliotic intervertebral joint and its incorporation into scoliotic musculoskeletal spine models.

Integrating structural variant calling, annotation and prioritization into whole genome analysis workflows: a practical application in the molecular diagnosis of neurodevelopmental disorders

Background: WGS is increasingly used as a first-line diagnostic test for patients with rare genetic diseases such as neurodevelopmental disorders (NDD). Clinical applications require a robust infrastructure to support processing, storage and analysis of WGS data. The identification and interpretation of SVs from WGS data also needs to be improved. Finally, there is a need for a prioritization system that enables downstream clinical analysis and facilitates data interpretation. Here, we present the results of a clinical application of WGS in a cohort of patients with NDD. Methods: We developed highly portable workflows for processing WGS data, including alignment, quality control, and variant calling of SNVs and SVs. A benchmark analysis of state-of-the-art SV detection tools was performed to select the most accurate combination for SV calling. A gene-based prioritization system was also implemented to support variant interpretation. Results: Using a benchmark analysis, we selected the most accurate combination of tools to improve SV detection from WGS data and build a dedicated pipeline. Our workflows were used to process WGS data from 77 NDD patient-parent families. The prioritization system supported downstream analysis and enabled molecular diagnosis in 32% of patients, 25% of which were SVs and suggested a potential diagnosis in 20% of patients, requiring further investigation to achieve diagnostic certainty. Conclusion: Our data suggest that the integration of SNVs and SVs is a main factor that increases diagnostic yield by WGS and show that the adoption of a dedicated pipeline improves the process of variant detection and interpretation.

Supervised and weakly supervised counting-by-segmentation: the fluorescent microscopy use case

This thesis focuses on automating the time-consuming task of manually counting activated neurons in fluorescent microscopy images, which is used to study the mechanisms underlying torpor. The traditional method of manual annotation can introduce bias and delay the outcome of experiments, so the author investigates a deep-learning-based procedure to automatize this task. The author explores two of the main convolutional-neural-network (CNNs) state-of-the-art architectures: UNet and ResUnet family model, and uses a counting-by-segmentation strategy to provide a justification of the objects considered during the counting process. The author also explores a weakly-supervised learning strategy that exploits only dot annotations. The author quantifies the advantages in terms of data reduction and counting performance boost obtainable with a transfer-learning approach and, specifically, a fine-tuning procedure. The author released the dataset used for the supervised use case and all the pre-training models, and designed a web application to share both the counting process pipeline developed in this work and the models pre-trained on the dataset analyzed in this work.

MR in vivo tractography for the reconstruction of cranial nerves course

Aim The aim of my Ph.D. was to implement a diffusion tensor tractography (DTT) pipeline to reconstruct cranial nerve I (olfactory) to study COVID-19 patients, and anterior optic pathway (AOP, including optic nerve, chiasm, and optic tract) to study patients with sellar/parasellar tumors, and with Leber’s Hereditary Optic Neuropathy (LHON). Methods We recruited 23 patients with olfactory dysfunction after COVID-19 infection (mean age 37±14 years, 12 females); 27 patients with sellar/parasellar tumors displacing the optic chiasm eligible for endonasal endoscopic surgery (mean age 53. ±16.4 years, 13 female) and 6 LHON patients (mutation 11778/MT-ND4, mean age 24.9±15.7 years). Sex- and age-matched healthy control were also recruited. In LHON patients, optical coherence tomography (OCT) was performed. Acquisitions were performed on a clinical high field 3-T MRI scanner, using a multi-shell HARDI (High Angular Resolution Diffusion Imaging) sequence (b-values 0-300-1000-2000 s/mm2, 64 maximum gradient directions, 2mm3 isotropic voxel). DTT was performed with a multi-tissue spherical deconvolution approach and mean diffusivity (MD) DTT metrics were compared with healthy controls using an unpaired t-test. Correlations of DTT metrics with clinical data were sought by regression analysis. Results In all 23 hypo/anosmic patients with previous COVID-19 infection the CN I was successfully reconstructed with no DTT metrics alterations, thus suggesting the pathogenetic role of central olfactory cortical system dysfunction. In all 27 patients with sellar/parasellar tumors the AOP was reconstructed, and in 11/13 (84.7%) undergoing endonasal endoscopic surgery the anatomical fidelity of the reconstruction was confirmed; a significant decrease in MD within the chiasma (p<0.0001) was also found. In LHON patients a reduction of MD in the AOP was significantly associated with OCT parameters (p=0.036). Conclusions Multi-shell HARDI diffusion-weighted MRI followed by multi-tissue spherical deconvolution for the DTT reconstruction of the CN I and AOP has been implemented, and its utility demonstrated in clinical practice.

Development of unsupervised learning methods with applications to life sciences data

Machine Learning makes computers capable of performing tasks typically requiring human intelligence. A domain where it is having a considerable impact is the life sciences, allowing to devise new biological analysis protocols, develop patients’ treatments efficiently and faster, and reduce healthcare costs. This Thesis work presents new Machine Learning methods and pipelines for the life sciences focusing on the unsupervised field. At a methodological level, two methods are presented. The first is an “Ab Initio Local Principal Path” and it is a revised and improved version of a pre-existing algorithm in the manifold learning realm. The second contribution is an improvement over the Import Vector Domain Description (one-class learning) through the Kullback-Leibler divergence. It hybridizes kernel methods to Deep Learning obtaining a scalable solution, an improved probabilistic model, and state-of-the-art performances. Both methods are tested through several experiments, with a central focus on their relevance in life sciences. Results show that they improve the performances achieved by their previous versions. At the applicative level, two pipelines are presented. The first one is for the analysis of RNA-Seq datasets, both transcriptomic and single-cell data, and is aimed at identifying genes that may be involved in biological processes (e.g., the transition of tissues from normal to cancer). In this project, an R package is released on CRAN to make the pipeline accessible to the bioinformatic Community through high-level APIs. The second pipeline is in the drug discovery domain and is useful for identifying druggable pockets, namely regions of a protein with a high probability of accepting a small molecule (a drug). Both these pipelines achieve remarkable results. Lastly, a detour application is developed to identify the strengths/limitations of the “Principal Path” algorithm by analyzing Convolutional Neural Networks induced vector spaces. This application is conducted in the music and visual arts domains.

Unveiling the expansion history of the universe with cosmic chronometers and gravitational waves

This Thesis explores two novel and independent cosmological probes, Cosmic Chronometers (CCs) and Gravitational Waves (GWs), to measure the expansion history of the Universe. CCs provide direct and cosmology-independent measurements of the Hubble parameter H(z) up to z∼2. In parallel, GWs provide a direct measurement of the luminosity distance without requiring additional calibration, thus yielding a direct measurement of the Hubble constant H0=H(z=0). This Thesis extends the methodologies of both of these probes to maximize their scientific yield. This is achieved by accounting for the interplay of cosmological and astrophysical parameters to derive them jointly, study possible degeneracies, and eventually minimize potential systematic effects. As a legacy value, this work also provides interesting insights into galaxy evolution and compact binary population properties. The first part presents a detailed study of intermediate-redshift passive galaxies as CCs, with a focus on the selection process and the study of their stellar population properties using specific spectral features. From their differential aging, we derive a new measurement of the Hubble parameter H(z) and thoroughly assess potential systematics. In the second part, we develop a novel methodology and pipeline to obtain joint cosmological and astrophysical population constraints using GWs in combination with galaxy catalogs. This is applied to GW170817 to obtain a measurement of H0. We then perform realistic forecasts to predict joint cosmological and astrophysical constraints from black hole binary mergers for upcoming gravitational wave observatories and galaxy surveys. Using these two probes we provide an independent reconstruction of H(z) with direct measurements of H0 from GWs and H(z) up to z∼2 from CCs and demonstrate that they can be powerful independent probes to unveil the expansion history of the Universe.

Quantitative susceptibility mapping as biomarker of neurodegeneration: methodological models and clinical applications

Quantitative Susceptibility Mapping (QSM) is an advanced magnetic resonance technique that can quantify in vivo biomarkers of pathology, such as alteration in iron and myelin concentration. It allows for the comparison of magnetic susceptibility properties within and between different subject groups. In this thesis, QSM acquisition and processing pipeline are discussed, together with clinical and methodological applications of QSM to neurodegeneration. In designing the studies, significant emphasis was placed on results reproducibility and interpretability. The first project focuses on the investigation of cortical regions in amyotrophic lateral sclerosis. By examining various histogram susceptibility properties, a pattern of increased iron content was revealed in patients with amyotrophic lateral sclerosis compared to controls and other neurodegenerative disorders. Moreover, there was a correlation between susceptibility and upper motor neuron impairment, particularly in patients experiencing rapid disease progression. Similarly, in the second application, QSM was used to examine cortical and sub-cortical areas in individuals with myotonic dystrophy type 1. The thalamus and brainstem were identified as structures of interest, with relevant correlations with clinical and laboratory data such as neurological evaluation and sleep records. In the third project, a robust pipeline for assessing radiomic susceptibility-based features reliability was implemented within a cohort of patients with multiple sclerosis and healthy controls. Lastly, a deep learning super-resolution model was applied to QSM images of healthy controls. The employed model demonstrated excellent generalization abilities and outperformed traditional up-sampling methods, without requiring a customized re-training. Across the three disorders investigated, it was evident that QSM is capable of distinguishing between patient groups and healthy controls while establishing correlations between imaging measurements and clinical data. These studies lay the foundation for future research, with the ultimate goal of achieving earlier and less invasive diagnoses of neurodegenerative disorders within the context of personalized medicine.

Models and methods for damage and quantitative risk assessment of Natech scenarios

Natural events are a widely recognized hazard for industrial sites where relevant quantities of hazardous substances are handled, due to the possible generation of cascading events resulting in severe technological accidents (Natech scenarios). Natural events may damage storage and process equipment containing hazardous substances, that may be released leading to major accident scenarios called Natech events. The need to assess the risk associated with Natech scenarios is growing and methodologies were developed to allow the quantification of Natech risk, considering both point sources and linear sources as pipelines. A key element of these procedures is the use of vulnerability models providing an estimation of the damage probability of equipment or pipeline segment as a result of the impact of the natural event. Therefore, the first aim of the PhD project was to outline the state of the art of vulnerability models for equipment and pipelines subject to natural events such as floods, earthquakes, and wind. Moreover, the present PhD project also aimed at the development of new vulnerability models in order to fill some gaps in literature. In particular, a vulnerability model for vertical equipment subject to wind and to flood were developed. Finally, in order to improve the calculation of Natech risk for linear sources an original methodology was developed for Natech quantitative risk assessment methodology for pipelines subject to earthquakes. Overall, the results obtained are a step forward in the quantitative risk assessment of Natech accidents. The tools developed open the way to the inclusion of new equipment in the analysis of Natech events, and the methodology for the assessment of linear risk sources as pipelines provides an important tool for a more accurate and comprehensive assessment of Natech risk.

Network analysis and machine learning assist drug repurposing and safety assessment in neurological diseases

In recent decades, two prominent trends have influenced the data modeling field, namely network analysis and machine learning. This thesis explores the practical applications of these techniques within the domain of drug research, unveiling their multifaceted potential for advancing our comprehension of complex biological systems. The research undertaken during this PhD program is situated at the intersection of network theory, computational methods, and drug research. Across six projects presented herein, there is a gradual increase in model complexity. These projects traverse a diverse range of topics, with a specific emphasis on drug repurposing and safety in the context of neurological diseases. The aim of these projects is to leverage existing biomedical knowledge to develop innovative approaches that bolster drug research. The investigations have produced practical solutions, not only providing insights into the intricacies of biological systems, but also allowing the creation of valuable tools for their analysis. In short, the achievements are: • A novel computational algorithm to identify adverse events specific to fixed-dose drug combinations. • A web application that tracks the clinical drug research response to SARS-CoV-2. • A Python package for differential gene expression analysis and the identification of key regulatory "switch genes". • The identification of pivotal events causing drug-induced impulse control disorders linked to specific medications. • An automated pipeline for discovering potential drug repurposing opportunities. • The creation of a comprehensive knowledge graph and development of a graph machine learning model for predictions. Collectively, these projects illustrate diverse applications of data science and network-based methodologies, highlighting the profound impact they can have in supporting drug research activities.

Developing a radiomic based machine learning model to identify patients with resected non-small-cell lung cancer at high risk of relapse

Background There is a wide variation of recurrence risk of Non-small-cell lung cancer (NSCLC) within the same Tumor Node Metastasis (TNM) stage, suggesting that other parameters are involved in determining this probability. Radiomics allows extraction of quantitative information from images that can be used for clinical purposes. The primary objective of this study is to develop a radiomic prognostic model that predicts a 3 year disease free-survival (DFS) of resected Early Stage (ES) NSCLC patients. Material and Methods 56 pre-surgery non contrast Computed Tomography (CT) scans were retrieved from the PACS of our institution and anonymized. Then they were automatically segmented with an open access deep learning pipeline and reviewed by an experienced radiologist to obtain 3D masks of the NSCLC. Images and masks underwent to resampling normalization and discretization. From the masks hundreds Radiomic Features (RF) were extracted using Py-Radiomics. Hence, RF were reduced to select the most representative features. The remaining RF were used in combination with Clinical parameters to build a DFS prediction model using Leave-one-out cross-validation (LOOCV) with Random Forest. Results and Conclusion A poor agreement between the radiologist and the automatic segmentation algorithm (DICE score of 0.37) was found. Therefore, another experienced radiologist manually segmented the lesions and only stable and reproducible RF were kept. 50 RF demonstrated a high correlation with the DFS but only one was confirmed when clinicopathological covariates were added: Busyness a Neighbouring Gray Tone Difference Matrix (HR 9.610). 16 clinical variables (which comprised TNM) were used to build the LOOCV model demonstrating a higher Area Under the Curve (AUC) when RF were included in the analysis (0.67 vs 0.60) but the difference was not statistically significant (p=0,5147).

Assessing model credibility of CT-Based finite element model to predict bone strength

Osteoporosis is one of the major causes of mortality among the elderly. Nowadays, areal bone mineral density (aBMD) is used as diagnostic criteria for osteoporosis; however, this is a moderate predictor of the femur fracture risk and does not capture the effect of some anatomical and physiological properties on the bone strength estimation. Data from past research suggest that most fragility femur fractures occur in patients with aBMD values outside the pathological range. Subject-specific finite element models derived from computed tomography data are considered better tools to non-invasively assess hip fracture risk. In particular, the Bologna Biomechanical Computed Tomography (BBCT) is an In Silico methodology that uses a subject specific FE model to predict bone strength. Different studies demonstrated that the modeling pipeline can increase predictive accuracy of osteoporosis detection and assess the efficacy of new antiresorptive drugs. However, one critical aspect that must be properly addressed before using the technology in the clinical practice, is the assessment of the model credibility. The aim of this study was to define and perform verification and uncertainty quantification analyses on the BBCT methodology following the risk-based credibility assessment framework recently proposed in the VV-40 standard. The analyses focused on the main verification tests used in computational solid mechanics: force and moment equilibrium check, mesh convergence analyses, mesh quality metrics study, evaluation of the uncertainties associated to the definition of the boundary conditions and material properties mapping. Results of these analyses showed that the FE model is correctly implemented and solved. The operation that mostly affect the model results is the material properties mapping step. This work represents an important step that, together with the ongoing clinical validation activities, will contribute to demonstrate the credibility of the BBCT methodology.

Machine learning "as a service" for high energy physics (MLaaS4HEP): evolution of a framework for ML-based physics analyses

The scientific success of the LHC experiments at CERN highly depends on the availability of computing resources which efficiently store, process, and analyse the amount of data collected every year. This is ensured by the Worldwide LHC Computing Grid infrastructure that connect computing centres distributed all over the world with high performance network. LHC has an ambitious experimental program for the coming years, which includes large investments and improvements both for the hardware of the detectors and for the software and computing systems, in order to deal with the huge increase in the event rate expected from the High Luminosity LHC (HL-LHC) phase and consequently with the huge amount of data that will be produced. Since few years the role of Artificial Intelligence has become relevant in the High Energy Physics (HEP) world. Machine Learning (ML) and Deep Learning algorithms have been successfully used in many areas of HEP, like online and offline reconstruction programs, detector simulation, object reconstruction, identification, Monte Carlo generation, and surely they will be crucial in the HL-LHC phase. This thesis aims at contributing to a CMS R&D project, regarding a ML "as a Service" solution for HEP needs (MLaaS4HEP). It consists in a data-service able to perform an entire ML pipeline (in terms of reading data, processing data, training ML models, serving predictions) in a completely model-agnostic fashion, directly using ROOT files of arbitrary size from local or distributed data sources. This framework has been updated adding new features in the data preprocessing phase, allowing more flexibility to the user. Since the MLaaS4HEP framework is experiment agnostic, the ATLAS Higgs Boson ML challenge has been chosen as physics use case, with the aim to test MLaaS4HEP and the contribution done with this work.

TARO: Infrastruttura per il Confronto di Testate Giornalistiche Internazionali

TARO (Tons of Articles Ready to Outline) è un progetto che ha come scopo quello di realizzare un sistema per la raccolta, l'analisi e il confronto di articoli di giornali online. Sono state scelte come fonti testate giornalistiche internazionali e i loro canali di pubblicazione, come ad esempio i Feed RSS e le Homepage. Gli articoli vengono quindi analizzati attraverso NER e Sentiment Analysis per poi individuare quali argomenti siano trattati da più testate e quali invece risultino esclusivi di una sola, sfruttando algoritmi di similarità. Il progetto è sviluppato in Python e sono utilizzate diverse librerie, tra cui Scrapy, per la raccolta di articoli, Argos, per la traduzione delle notizie al fine di allinearle linguisticamente, SpaCy, per le analisi semantiche, e Pandas per la visualizzazione dei risultati ottenuti. Uno degli obiettivi è sfruttare questa pipeline al fine di effettuare analisi socio-culturali interessanti utilizzando le informazioni date dagli articoli giornalistici stessi, oltre che osservare le potenzialità delle analisi semantiche fatte su notiziari.

Machine Learning in ambito assicurativo Un caso di studio

Il Machine Learning si sta rivelando una tecnologia dalle incredibili potenzialità nei settori più disparati. Le diverse tecniche e gli algoritmi che vi fanno capo abilitano analisi dei dati molto più efficaci rispetto al passato. Anche l’industria assicurativa sta sperimentando l’adozione di soluzioni di Machine Learning e diverse sono le direzioni di innovamento che ne stanno conseguendo, dall’efficientamento dei processi interni all’offerta di prodotti rispondenti in maniera adattiva alle esigenze del cliente. Questo lavoro di tesi è stato realizzato durante un tirocinio presso Unisalute S.p.A., la prima assicurazione in ambito sanitario in Italia. La criticità intercettata è stata la sovrastima del capitale da destinare a riserva a fronte dell’impegno nei confronti dell’assicurato: questo capitale immobilizzato va a sottrarre risorse ad investimenti più proficui nel medio e lungo termine, per cui è di valore stimarlo appropriatamente. All'interno del settore IT di Unisalute, ho lavorato alla progettazione e implementazione di un modello di Machine Learning che riesca a prevedere se un sinistro appena preso in gestione sarà liquidato o meno. Dotare gli uffici impegnati nella determinazione del riservato di questa stima aggiuntiva basata sui dati, sarebbe di notevole supporto. La progettazione del modello di Machine Learning si è articolata in una Data Pipeline contenente le metodologie più efficienti con riferimento al preprocessamento e alla modellazione dei dati. L’implementazione ha visto Python come linguaggio di programmazione; il dataset, ottenuto a seguito di estrazioni e integrazioni a partire da diversi database Oracle, presenta una cardinalità di oltre 4 milioni di istanze caratterizzate da 32 variabili. A valle del tuning degli iperparamentri e dei vari addestramenti, si è raggiunta un’accuratezza dell’86% che, nel dominio di specie, è ritenuta più che soddisfacente e sono emersi contributi non noti alla liquidabilità dei sinistri.