UQ4ML | COMETA Workshop on Uncertainty Quantification for Machine Learning

15 Sept 2025, 14:00 → 18 Sept 2025, 22:15 Europe/Paris

Amphithéâtre Claude Bloch (IPhT) (CEA Paris-Saclay)

Alessandra Cappati (Universite Catholique de Louvain (UCL) (BE)), Claudius Krause (HEPHY Vienna (ÖAW)), Harold Erbin, Jean-Baptiste Blanchard (CEA/IRFU,Centre d'etude de Saclay Gif-sur-Yvette (FR)), Karolos Potamianos (University of Warwick (GB)), Marco Letizia, Riccardo Finotello (CEA Paris-Saclay), Shamik Ghosh (Centre National de la Recherche Scientifique (FR))

UQ4ML | COMETA Workshop on Uncertainty Quantification for Machine Learning

 

---

Mathematicians and physicists perspectives

In recent years, the landscape of scientific research has been dramatically reshaped by advancements in AI and machine learning (ML). These tools have enabled us to process vast amounts of data and uncover complex patterns with unprecedented efficiency. However, as we delve deeper into these realms, the importance of understanding and quantifying uncertainty in our computations has become increasingly apparent. Uncertainty Quantification (UQ) is not just about acknowledging the limitations of our models and data, but also about harnessing this understanding to make more robust predictions and get to more reliable conclusions. It is a vital component of scientific rigour, enabling us to navigate the complexities of real-world systems with confidence.

Objective and Goals

We aim to foster a cross-disciplinary dialogue on the challenges and opportunities presented by UQ: a topic developed across decades by applied mathematicians, and used by a wide variety of scientists for many applications, High Energy Physics (HEP) being only one particular case. We will discuss how results from different research fields can enhance our understanding in physics, improve the reliability of mathematical models, guide the development of scientific AI and ML tools in the future.

Organizing Committee

The French Atomic and Alternative Energy Commission (CEA), backed by excellence in fundamental research, provides concrete solutions to society's needs for low carbon energy, numerical challenges, and technologies for future medicine. The Laboratory of Artificial Intelligence and Data Science (LIAD) of CEA is dedicated to tackle the challenges of the ever-growing need for precise, trustworthy, and robust solutions in statistics and machine learning.

The Institute of Theoretical Physics (IPhT), known for its excellency and devoted to deal with fundamental research, it also provides capable solutions for modern applications, such as quantum computing.

Keynote Speakers

Aurore Lomet	CEA Paris-Saclay
Giuliano Panico	University of Florence
Christian Glaser	Uppsala University
Anja Butter	ITP Heidelberg, LPNHE Paris
Merlin Keller	EDF Paris
Jorge Fernandez-de-Cossio-Diaz	IPhT Paris-Saclay

 

---

 

For security reasons, the number of participants will be limited to 40.

Monday 15 September

Organisation: Welcome

Convener: Dr Riccardo Finotello (CEA Paris-Saclay)

1 Welcome to UQ4ML

Speakers: Alessandra Cappati (Universite Catholique de Louvain (UCL) (BE)), Dr Marco Letizia, Dr Riccardo Finotello (CEA Paris-Saclay), Shamik Ghosh (Centre National de la Recherche Scientifique (FR)), Karolos Potamianos (University of Warwick (GB)), Dr Claudius Krause (HEPHY Vienna (ÖAW))

20250915_uq4ml_welcome.pdf

Simulations and Coding: Keynote

Convener: Jean-Baptiste Blanchard (CEA/IRFU,Centre d'etude de Saclay Gif-sur-Yvette (FR))

2 Quantifying uncertainties in computer models: an overview

Predictions from computer models are now extensively in industrial studies, to complement, or even sometimes replace, field experiments. Such numerical experiments have key advantages, such as reduced costs, and added flexibility. However, they raise the question of assessing the validity of computer model predictions, with respect to the physical phenomena they seek to reproduce. This is the goal of verification, validation and uncertainty quantification (VVUQ), a process whose development is an active field of research in applied mathematics. The goal of this talk is to present the different objectives and challenges of VVUQ, as well as a generic workflow that can be applied to virtually any uncertainty quantification problem. We then discuss software libraries that implement this methodology, as well as open problems and perspectives.

Speaker: Merlin Keller (EDF)

20250915_uq_computer_models.pdf

15:15 Coffee break

Simulations and Coding: Talks

Convener: Jean-Baptiste Blanchard (CEA/IRFU,Centre d'etude de Saclay Gif-sur-Yvette (FR))

3 Uncertainty in AI-driven Monte Carlo simulations

In the study of complex systems, evaluating physical observables often requires sampling representative configurations via Monte Carlo techniques. These methods rely on repeated evaluations of the system's energy and force fields, which can become computationally expensive. To accelerate these simulations, deep learning models are increasingly employed as surrogate functions to approximate the energy landscape or force fields. However, such models introduce epistemic uncertainty in their predictions, which may propagate through the sampling process and affect the system's macroscopic behavior. In our work, we present the Penalty Ensemble Method (PEM) to quantify epistemic uncertainty and mitigate its impact on Monte Carlo sampling. Our approach introduces an uncertainty-aware modification of the Metropolis acceptance rule, which increases the rejection probability in regions of high uncertainty, thereby enhancing the reliability of the simulation outcomes.

Speaker: Dimitrios Tzivrailis (CEA Paris-Saclay)

20250915_uncertainty_ai_simulations.pdf

4 Parameter Estimation with Neural Simulation-Based Inference in ATLAS

Neural Simulation-Based Inference (NSBI) is a powerful class of machine learning (ML)-based methods for statistical inference that naturally handle high dimensional parameter estimation without the need to bin data into low-dimensional summary histograms. Such methods are promising for a range of measurements at the Large Hadron Collider, where no single observable may be optimal to scan over the entire theoretical phase space under consideration, or where binning data into histograms could result in a loss of sensitivity. This work develops an NSBI framework that, for the first time, allows NSBI to be applied to a full-scale LHC analysis, by successfully incorporating a large number of systematic uncertainties, quantifying the uncertainty coming from finite training statistics, developing a method to construct confidence intervals, and demonstrating a series of intermediate diagnostic checks that can be performed to validate the robustness of the method. As an example, the power and feasibility of the method are demonstrated for an off-shell Higgs boson couplings measurement in the four lepton decay channel, using ATLAS experiment simulated samples. The proposed method is a generalisation of the standard statistical framework at the LHC, and can benefit a large number of physics analyses. This work serves as a blueprint for measurements at the LHC using NSBI.

This talk covers https://arxiv.org/abs/2412.01600 which is a methodology paper detailing the NSBI technique used in the physics paper https://arxiv.org/abs/2412.01548 .

Speaker: David Rousseau (IJCLab-Orsay)

20250915_atlas_nsbi2.pdf

Tuesday 16 September

HEP - Theory: Keynote

Convener: Dr Marco Letizia

5 Machine learning for optimal BSM sensitivity

I will review the parametrized classifiers for optimizing the sensitivity to EFT operators and some the machine-learning approaches for general anomaly detection. Particular attention will be devoted to validation procedures and ways to treat uncertainties.

Speaker: Giuliano Panico (University of Florence and INFN Florence)

20250916_ml_bsm_sensitivity.pdf

10:30 Coffee break

HEP - Theory: Talks

Convener: Dr Marco Letizia

6 Determination and validation of modeling and theory uncertaintie

I discuss how uncertainties related to machine learning modeling of a regression problem, as well as those related to missing theoretical information, can be estimated and subsequently validated. Even though these uncertainties are intrinsically Bayesian, given that there is only one underlying true theory and true model, they can be determined both in a Bayesian and frequentist framework. I show how this can be done in the context of the determination of the parton distributions that encode the structure of the proton. I further show how results can be validated by means of closure tests.

Speaker: Stefano Forte (Università degli Studi e INFN Milano (IT))

uncval.pdf

7 Parton Distributions from Neural Networks: Analytical Results

Parton Distribution Functions (PDFs) play a crucial role in describing experimental data at hadron colliders and provide insight into proton structure. As the LHC enters an era of high-precision measurements, a robust PDF determination with a reliable uncertainty quantification has become increasingly important to match the experimental precision. The NNPDF collaboration has pioneered the use of Machine Learning (ML) techniques for PDF determination. In this work, we develop a theoretical framework based on the Neural Tangent Kernel (NTK) to analyse the training dynamics of Neural Networks. This approach allows us to derive, under certain assumptions, an analytical description of how the neural network evolves during training, enabling us to better understand the NNPDF methodology and its dependence on the underlying model architecture. Notably, we demonstrate that our results contrast, to some extent, with the standard picture of the lazy training regime commonly discussed in the ML community.

Speaker: Amedeo Chiefa

20250916_nn_parton_distributions.pdf

8 Energy Flow Networks for Jet Quenching Studies

The phenomena of Jet Quenching, a key signature of the Quark-Gluon Plasma (QGP) formed in Heavy-Ion (HI) collisions, provides a window of insight into the properties of the primordial liquid. In this study, we evaluate the discriminating power of Energy Flow Networks (EFNs), enhanced with substructure observables, in distinguishing between jets stemming from proton-proton (pp) and jets stemming from HI collisions. This work is a crucial step towards separating HI jets that were quenched from those with little or no modification by the interaction with the QGP on a jet-by-jet basis. We trained simple Energy Flow Networks (EFNs) and further enhanced them by incorporating jet observables such as N-Subjettiness and Energy Flow Polynomials (EFPs). Our primary objective is to assess the effectiveness of these approaches in the context of Jet Quenching, exploring new phenomenological avenues by combining these models with various encodings of jet information. Initial evaluations using Linear Discriminant Analysis (LDA) set a performance baseline, which is significantly enhanced through simple Deep Neural Networks (DNNs), capable of capturing non-linear relations expected in the data. Integrating both EFPs and N-Subjettiness observables into EFNs results in the most performant model over this task, achieving state-of-the-art ROC AUC values of approximately 0.84. This significant performance is noteworthy given that both medium response and underlying event contamination effects on the jet are taken into account. These results underscore the potential of combining EFNs with jet substructure observables to advance Jet Quenching studies and adjacent areas, paving the way for deeper insights into the properties of the QGP. Results on a variation of EFNs, Moment EFNs (MEFNs), which can achieve comparable performance with a more manageable and, in turn interpretable, latent space, will be presented.

Speaker: João A. Gonçalves (LIP - IST)

UQ4ML.pdf

12:30 Lunch break

HEP - Experiment: Keynote

Convener: Shamik Ghosh (Centre National de la Recherche Scientifique (FR))

9 Uncertainty quantification for deep learning in astroparticle physics

In this contribution I will review the use cases of uncertainty quantification with deep learning in high-energy astroparticle physics. Among other things, I will present the combination of neural networks with conditional normalizing flows to predict the Posterior for all quantities of interest. This Ansatz can be further expanded with the snowstorm method developed by the IceCube collaboration to include systematic uncertainties in the Posterior prediction by sampling from the systematic uncertainties during MC generation.

Speaker: Christian Glaser (Uppsala University)

talk_UQ in Astroparticlephysics_Glaser.pdf

15:00 Coffee break

HEP - Experiment: Talks

Convener: Shamik Ghosh (Centre National de la Recherche Scientifique (FR))

10 Interdisciplinary Digital Twin Engine InterTwin for calorimeter simulation

The interTwin project develops an open-source Digital Twin Engine to integrate application-specific Digital Twins (DTs) across scientific domains. Its framework for the development of DTs supports interoperability, performance, portability and accuracy. As part of this initiative, we implemented the CaloINN normalizing-flow model for calorimeter simulations within the interTwin framework. Calorimeter shower simulations are computationally expensive, and generative models offer an efficient alternative. However, achieving a balance between accuracy and speed remains a challenge, with distribution tail modeling being a key limitation. CaloINN provides a trade-off between simulation quality and efficiency. The ongoing study targets validating the model using high granularity simulations from the Open Data Detector, as well as introducing a set of post-processing modifications of analysis-level observables aimed at improving the accuracy of distribution tails.

Speaker: Vera Maiboroda (CNRS, IJCLab)

20250916_calorimeter_shower_simulations.pdf

11 Scaling laws for amplitude surrogates

Fast and precise evaluations of scattering amplitudes even in the case of precision calculations is essential for event generation tools at the HL-LHC. We explore the scaling behavior of the achievable precision of neural networks in this regression problem for multiple architectures, including a Lorentz symmetry aware multilayer perceptron and the L-GATr architecture. L-GATr is equivariant with respect to the Lorentz group by its internal embedding living in the geometric algebra defined by the flat space-time metric. This study addresses in particular the scaling behavior of uncertainty estimations using state of the art methods.

Speaker: Joaquin Iturriza Ramirez (Centre National de la Recherche Scientifique (FR))

20250916_scaling_laws_amplitude_surrogates.pdf

12 Fair Universe HiggsML Uncertainty Challenge

The Fair Universe project organised the HiggsML Uncertainty Challenge, which took place from 12th September 2024, to 14th March 2025. This groundbreaking competition in high-energy physics (HEP) and machine learning was the first to strongly emphasis on uncertainties, focusing on mastering both the uncertainties in the input training data and providing credible confidence intervals in the results.
The challenge revolved around measuring the Higgs to tau+ tau- cross-section, similar to the HiggsML challenge held on Kaggle in 2014, using a dataset representing the 4-momentum signal state. Participants were tasked with developing advanced analysis techniques capable of measuring the signal strength and generating confidence intervals that included both statistical and systematic uncertainties, such as those related to detector calibration and background levels. The accuracy of these intervals was automatically evaluated using pseudo-experiments to assess correct coverage.
Techniques that effectively managed the impact of systematic uncertainties were expected to perform best, contributing to the development of uncertainty-aware AI techniques for HEP and potentially other fields. The competition was hosted on Codabench, an evolution of the Codalab platform, and leveraged significant resources from the NERSC infrastructure to handle the thousands of required pseudo-experiments.
This competition was selected as a NeurIPS competition, and the preliminary results were presented at the NeurIPS 2024 conference in December. As the challenge concluded in March 2025, an account of the winning solutions will be presented at this workshop.

Speaker: Ragansu Chakkappai (IJCLab-Orsay)

Fair Universe_ HiggsML Uncertainty Challenge - UQ_COMETA.pdf

Fair Universe_ HiggsML Uncertainty Challenge - UQ_COMETA.pdf

Fair Universe_ HiggsML Uncertainty Challenge - UQ_COMETA.pptx

Wednesday 17 September

Data analysis, Time Series, Causal analysis: Keynote

Convener: Riccardo Finotello (CEA Paris-Saclay)

13 Introduction to Causality for Time Series and HSIC

Causality, in Pearl’s framework, is defined through structural causal models: systems of structural equations with exogenous variables and a directed acyclic graph that encodes cause–effect relations. In contrast, correlation, which often forms the basis of artificial intelligence models, quantifies statistical association and may arise from confounding or indirect paths without implying a directed effect.
In physics, causal analysis enables the identification of mechanisms that support predictive models under interventions. Correlation alone cannot determine whether an observed dependency originates from a physical mechanism, a hidden common cause, or a statistical artifact. Moreover, consideration of causal structure provides a basis for interpretable artificial intelligence models, since the reasoning aligns with the formal description of interactions in physics.
In this context, causal discovery aims to recover the underlying causal graph from data. Approaches based on independence testing require measures that remain valid under realistic assumptions. Linear methods such as partial correlation impose Gaussianity and linear relations. Kernel-based measures such as the Hilbert–Schmidt Independence Criterion (HSIC) relax these constraints and detect nonlinear dependencies. Building on this foundation, this presentation introduces a causal discovery method for time series, such as sensor data, that employs kernel dependence measures, integrates kernel approximations to improve computational efficiency, and takes into account the temporal dependence in the HSIC to reduce bias from autocorrelation. The results indicate that the proposed method achieves consistent performances across different levels of dependence and length of time series compared to alternative approaches.

Speaker: Aurore Lomet (CEA Paris-Saclay)

20250917_causality_hsic.pdf

10:30 Coffee break

Data analysis, Time Series, Causal analysis: Talks

Convener: Riccardo Finotello (CEA Paris-Saclay)

14 Calibrated and uncertain? Evaluating uncertainty estimates in binary classification models

Rigorous statistical methods, including the estimation of parameter values and their uncertainties, underpins the validity of scientific discovery, and has been especially important in the natural sciences. In the age of data-driven modeling, where the complexity of data and statistical models grow exponentially as computing power increases, uncertainty quantification has become exceedingly difficult and a plethora of techniques using a wide variety of more or less developed mathematical foundations have been proposed.
In this case study we use the unifying theoretical framework of (approximate) nonparametric Bayesian inference and empirical tests on carefully created synthetic datasets to investigate qualitative properties of 6 different probabilistic machine learning algorithms for class probability and uncertainty estimation: (i) a neural network ensemble, (ii) neural network ensemble with conflictual loss, (iii) evidential deep learning, (iv) a single neural network with Monte Carlo Dropout, (v) Gaussian process classification and (vi) a Dirichlet process mixture model. We check if the algorithms produce results which reflect commonly desired statistical properties of uncertainty estimates in these kind of models such as calibration and an increase in uncertainty for out of distribution data points. Our results indicate that all algorithms are well calibrated, but none of the deep learning based algorithms provide uncertainties that reliably reflect lack of experimental evidence for out of distribution data points. We hope our study may serve as a clarifying example for researchers in the natural sciences trying to navigate the field of uncertainty quantification and especially to those developing new methods of uncertainty estimation for scientific data-driven modeling.

Speaker: Aurora Singstad Grefsrud (Western Norway University of Applied Sciences (NO))

Calibrated and uncertain (1).pdf

15 Unsupervised Anomaly Detection in Multivariate Time Series Using Public Benchmarks and Synthetic Data from Lorenzetti

Anomaly detection in multivariate time series is an important problem across various fields such as healthcare, financial services, manufacturing or physics detector monitoring. Accurately identifying the instances when defects occur is essential but challenging, as the types of anomalies are unknown beforehand and reliably labelled data are scarce.
We evaluate unsupervised transformer-based models and benchmark their performance against traditional methods on public data.
Furthermore, to address the lack of reliable labels, we use the Lorenzetti Shower simulator - a general-purpose framework for simulating high-energy calorimeters - where we introduce artificial defects to evaluate the sensitivity of various detection methods.

Speaker: Laura Boggia (Centre National de la Recherche Scientifique (FR))

202509_TSAD_LauraBoggia.pdf

16 Bayesian continual learning and forgetting in neural networks

Biological synapses effortlessly balance memory retention and flexibility, yet artificial neural networks still struggle with the extremes of catastrophic forgetting and catastrophic remembering. Here, we introduce Metaplasticity from Synaptic Uncertainty (MESU), a Bayesian framework that updates network parameters according to their uncertainty. This approach allows a principled combination of learning and forgetting that ensures that critical knowledge is preserved while unused or outdated information is gradually forgotten. Unlike standard Bayesian approaches – which risk becoming overly constrained because their posterior variances keep shrinking as evidence from all past tasks accumulates, and popular synaptic-consolidation-based continual-learning methods that rely on explicit task boundaries, MESU seamlessly adapts to streaming data. It further provides reliable epistemic uncertainty estimates, allowing out-of-distribution detection, the only computational cost being to sample the weights multiple times to provide proper output statistics. Experiments on image-classification benchmarks demonstrate that MESU mitigates catastrophic forgetting, while maintaining plasticity for new tasks. When training 200 sequential Permuted MNIST tasks, MESU outperforms established synaptic-consolidation-based continual learning techniques in terms of accuracy, capability to learn additional tasks, and out-of-distribution data detection. Additionally, due to its non-reliance on task boundaries, MESU outperforms conventional learning techniques on the incremental training of CIFAR-100 tasks consistently in a wide range of scenarios. Our results unify ideas from metaplasticity, Bayesian inference, and Hessian-based regularization, offering a biologically-inspired pathway to robust, perpetual learning.

Speaker: Kellian Cottart (Université Paris Saclay)

2504.13569v1.pdf

MESU_UQ_KellianCottart.pdf

12:30 Lunch break

Deep Learning and Uncertainty Quantification: Keynote

Convener: Alessandra Cappati (Universite Catholique de Louvain (UCL) (BE))

17 Uncertainty Quantification for Neural Networks in Particle Physics

Correctly calibrated uncertainties have always been a fundamental pillar of particle physics. As machine learning becomes increasingly integrated into both experimental and theoretical workflows, it is essential that neural network predictions include robust and reliable uncertainty estimates.

This talk will review current approaches to uncertainty estimation in neural networks, focusing on Bayesian neural networks, heteroscedastic loss functions, and repulsive ensembles. Their calibration and practical challenges will be discussed through examples from amplitude regression and unfolding. Additionally, we will explore how machine learning concepts of aleatoric and epistemic uncertainty relate to the statistical and systematic uncertainties familiar in particle physics.

Speaker: Anja Butter (Centre National de la Recherche Scientifique (FR))

COMETA_Saclay.pdf

15:00 Coffee break

Deep Learning and Uncertainty Quantification: Talks

Convener: Alessandra Cappati (Universite Catholique de Louvain (UCL) (BE))

18 Uncertainty Quantification in an ML Pattern Recognition Pipeline

Geometric learning pipelines have achieved state-of-the-art performance in High-Energy and Nuclear Physics reconstruction tasks like flavor tagging and particle tracking [1]. Starting from a point cloud of detector or particle-level measurements, a graph can be built where the measurements are nodes, and where the edges represent all possible physics relationships between the nodes. Depending on the size of the resulting input graph, a filtering stage may be needed to sparsify the graph connections. A Graph Neural Network will then build a latent representation of the input graph that can be used to predict, for example, whether two nodes (measurements) belong to the same particle or to classify a node as noise. The graph may then be partitioned into particle-level subgraphs, and a regression task used to infer the particle properties. Evaluating the uncertainty of the overall pipeline is important to measure and increase the statistical significance of the final result. How do we measure the uncertainty of the predictions of a multistep pattern recognition pipeline? How do we know which step of the pipeline contributes the most to the prediction uncertainty, and how do we distinguish between irreducible uncertainties arising from the aleatoric nature of our input data (detector noise, multiple scattering, etc) and epistemic uncertainties that we could reduce by using, for example, a larger model, or more training data?

We have developed an Uncertainty Quantification process for multistep pipelines to study these questions and applied it to the acorn particle tracking pipeline [2]. All our experiments are made using the TrackML open dataset [3]. Using the Monte Carlo Dropout method, we measure the data and model uncertainties of the pipeline steps, study how they propagate down the pipeline, and how they are impacted by the training dataset's size, the input data's geometry and physical properties. We will show that for our case study, as the training dataset grows, the overall uncertainty becomes dominated by aleatoric uncertainty, indicating that we had sufficient data to train the acorn model we chose to its full potential. We show that the ACORN pipeline yields high confidence in the track reconstruction and does not suffer from the miscalibration of the GNN model.

References:
[1] [2203.12852] Graph Neural Networks in Particle Physics: Implementations, Innovations, and Challenges
[2] acorn - GNN4ITkTeam
[3] Data - TrackML particle tracking challenge

Speaker: Lukas Péron

ACORN_UQ_UP_LukasPeron.pdf

arXiv paper

19 Uncertainty quantification for machine learning: a new approach for the Critical Heat Flux application

Critical Heat Flux (CHF) represents a concern for the nuclear safety, as it leads to a rapid drop down in the heat transfer between a heated surface and the liquid coolant in the core of nuclear reactors. This could cause several issues to the system, including structural damage and release of radioactive material.

The main challenge related to CHF prediction is the highly non-linear relationship with the physical features it depends on. For that reason, the prediction of CHF is often affected by large uncertainties.

In this research, the CHF database provided by the U.S. Nuclear Regulatory Commission (NRC) is utilized to develop machine learning (ML) methods for CHF prediction and robust uncertainty quantification (UQ) techniques. The performance of the ML models is assessed against established data-driven strategies, while a coverage-based approach is considered for UQ methods by using conformal prediction and a quality-driven loss function. It is found that it is possible to confidently estimate the prediction uncertainties with a 95% coverage regarding experimental CHF values.

Speaker: Michele Cazzola (CEA Paris-Saclay)

abstract.pdf

Thursday 18 September

Spotlight

Convener: Dr Harold Erbin

20 Machine learning of biological sequences

Over the last decade machine learning has had tremendous impact on biological sequence data analysis. In this talk, I will begin by introducing general issues related to biological sequence modeling. I will then review a selection of recent works on this topic, including: i) generative models for sequence design, ii) sampling of evolutionary paths between natural sequences of different classes, and iii) predictive models of directed evolution. I will also discuss some sources of uncertainty that arise with biological sequence data in different contexts (alignment, phylogenetic correlations, sampling noise, …), their potential impact on the models, and efforts to mitigate it.

Speaker: Jorge Fernández de Cossío Díaz (CEA Paris-Saclay)

10:00 Coffee break

Spotlight: Talks

Convener: Dr Harold Erbin

21 Interaction reconstruction in scintillator detectors for PET imaging - Deep Learning approach with uncertainty quantification

Positron Emission Tomography (PET) is a medical imaging modality that is powerful to follow biological processes. Nevertheless, it is of importance to increase its sensitivity to improve the contrast on the PET images and to decrease the patient exposure to radiation. One promising way is the use of the Time-of-Flight (ToF) of coincident gamma ray photons to get a more precise information of the annihilation location event by event. New instrumental developments are necessary to achieve this objective, especially ultra-fast detectors. The ClearMind project has developed a detection system based on a fast lead tungstate (PbWo4) monolithic scintillator detector and its conception necessitates new processing methods of its signals to reconstruct the gamma photons interactions.

This work focuses on the processing of the recorded signals to estimate the spatial coordinates of the gamma interactions within the detector. The complexity of these signals brings the necessity to use advanced tools and we have developed Deep Learning models, trained on simulation. We introduce a custom loss function that aims at estimating the inherent uncertainties due to the randomness of the signal formation and that incorporates the physical constraints of the detector. The results show the effectiveness of the proposed approach that provides a robust and reliable estimation of the interaction location. They highlight the benefit of the uncertainties estimation that will be exploited in the future for PET image reconstruction to discard or weight each individual event in the objective to improve the signal to noise ratio on the reconstructed image.

Speaker: Geoffrey Daniel

Organisation: Summary

Convener: Riccardo Finotello (CEA Paris-Saclay)

22 Wrap Up

Speakers: Dr Riccardo Finotello (CEA Paris-Saclay), Dr Claudius Krause (HEPHY Vienna (ÖAW)), Karolos Potamianos (University of Warwick (GB)), Shamik Ghosh (Centre National de la Recherche Scientifique (FR)), Dr Marco Letizia, Alessandra Cappati (Universite Catholique de Louvain (UCL) (BE))