International Workshop on Machine Learning ɑnd Quantum Computing Applications in Medicine ɑnd Physics
WMLQ2022
13 to 16 September 2022, Warsaw Poland
A welcome addess from Professor Krzysztof Kurek Director General of National Centre for Nuclear Research
Short communication from the workshop organizers.
Machine learning has evolved from an academic toolkit to a new computing technology that is already being used in technical and technology-related areas. The progress has been very impressive. While current academic research focuses on the improvement of core machine learning methods, the successful application of machine learning in real-life projects still requires constant monitoring and manual tuning.
This talk will highlight some of these lessons already learned, as well as some missing skills and points to consider in
machine learning projects.
Providing a real-time working system to localize the dangerous contaminant source is one of the main challenges for the city’s emergency response groups. Unfortunately, all proposed frameworks capable of estimating the contamination source localization based on recorded by the sensors network the substance concentrations cannot work in real-time. The reason is the significant computational time required by the applied dispersion models.The solution might be an application of the trained Artificial Neural Network (ANN) instead of the dispersion model in the reconstruction algorithm. To be used, the ANN must learn to simulate airborne contaminant transport. Training the ANN is computationally expensive, but once trained, the ANN would be a high-speed tool enabling the estimation of the contaminant concentration distribution.
This paper presents the results of training the ANN to predict the time evolution of the dispersion of the airborne contaminant over a city domain. The spatial distribution of the contaminant is the multidimensional function dependent on the weather conditions (wind direction and speed), coordinates of the contamination sources, the release rate, and its duration. Wwe try to answer what topology should be ANN to forecast the contaminant strength correctly at the given point of the urbanized area at a given time.
In recent years we could observe more text-to-image models, most notably DALLE, DALLE-2, and DALLE mini. Those models allow generating images based on user prompts. They were trained on datasets of images and captions crawled from the web. But those datasets contain some biases, especially those present in media. However DALLE and DALLE-2 are debiased models against at least race and gender, but DALLE-mini is not which makes him possible to pick up biases from the dataset. Some of those biases might be linked to misconceptions in society. Like prompt 'autistic child' will be 9 out of 9 cases a boy in preschool age, but 'autistic girl' will be presented as in secondary school years.
In the talk, we would like to show how not debiased models text-to-image could serve as a tool to study misconceptions about health that exists in society.
Firstly, I will enlighten how different quantum information is compared to classical information and how this can affect machine learning algorithms. Then I discuss different attempts to obtain quantum neural networks and their performances. In the last part I will focus on applications of machine learning to problems in quantum physics and medicine.
One of the possible applications of quantum computers in the near future are simulations of physics. An example are quantum gravitational systems associated with the Planck scale physics. Such systems are expected to be of the many-body type, which justifies utility of quantum computations in the analysis of their complex quantum behaviour. In this talk, loop quantum gravity - a leading candidate for the theory of quantum gravitational interactions - is considered. In this case, quantum geometry of space is represented by the so-called spin networks, i.e. graphs with nodes associated with the "atoms of space". A construction of quantum circuits which generate states of spin networks will be presented. Furthermore, a quantum algorithms which enable projection of states on physical subspace of Hilbert space and determination of amplitudes of transitions between different states of spin network are proposed. Results of implementation of the approach on IBM superconducting quantum computers will be presented. Obtained results provide building blocks for quantum simulations of complex spin networks, which can give insight into the Planck scale physics in the near future.
Support vector machines (SVMs) are a well-established classifier effectively deployed in an array of pattern recognition and classification tasks. In this work, we consider extending classic SVMs with quantum kernels and applying them to satellite data analysis.
The design and implementation of SVMs with quantum kernels (hybrid SVMs) is presented. It consists of the Quantum Kernel Estimation (QKE) procedure combined with a classic SVM training routine. The pixel data are mapped to the Hilbert space using ZZ-feature maps acting on the parameterized ansatz state. The parameters are optimized to maximize the kernel target alignment.
We approach the problem of cloud detection in satellite image data, which is one of the pivotal steps in both on-the-ground and on-board satellite image analysis processing chains. The experiments performed over the benchmark Landsat-8 multispectral dataset revealed that the simulated hybrid SVM successfully classifies satellite images with accuracy on par with classical SVMs.
We have witnessed the unprecedented success of deep learning in virtually all areas of science and industry, with medical image analysis not being an exception here. Although there are a plethora of deep learning-powered techniques that established the state of the art in the field, e.g., in the context of automatic delineation of human organs and tumors in various image modalities, deploying such methods in clinical settings is a challenging process. In this talk, we will show how deep learning, potentially coupled with computational fluid dynamics, can help uncover important clinical information to diagnose and monitor of the coronary artery disease from CCTA, or to analyze brain tumors from MRI. Also, we will discuss our approach for building Sens.AI – a CE marked deep learning product for automated brain tumor analysis. We will show how to design thorough evidence-based verification and validation procedures for such techniques in scenarios, in which collecting large, heterogeneous, and high-quality ground truth is time-consuming, user-dependent and error prone.
Poor air quality and its negative impact on health is currently one of the civilizational problems in Poland. The aim of this study was an attempt to verify and examine, on the basis of data on the number and causes of deaths registered in Bielański Hospital in Warsaw, the increase in the number of deaths in Poland in January 2017 recorded by Statistics Poland.
We analysed the data on the number and causes of deaths in the hospital from 2013 to 2018 using the methods of searching for anomalies and building models of the number of deaths depending on ambient temperature and air pollution levels.
We found that the increase in the number of deaths observed in the hospital in January 2017 was caused by respiratory system-related deaths. A model utilizing air temperature is not enough to explain the increase, but adding PM10 air pollution levels to the temperature model was sufficient to achieve this. Such a model attributes 8.3% of all deaths observed in January 2017 to air pollution.
Monte Carlo simulation of particle tracking in matter is the reference simulation method in the field of medical physics. It is heavily used in various applications such as
Monte Carlo simulation is a key tool both in academic research labs as well as industrial research and development services. Because of the very nature of the Monte Carlo method, involving iterative and stochastic estimation of numerous probability density functions, the computation time is high. In this presentation, we will review the recent use of Artificial Intelligence methods for Monte Carlo simulation in medical physics and their main associated challenges.
In PET medical imaging, the reconstruction of the spatial distribution
of the radiotracer in patient’s body is based on the photon pairs
grouped into time coincidences. Due to the limited resolution the selected
coincidences contain a fraction of events with a photon scattered in the
patient or detector material and photons accidentally registered in a coincidence.
Scatters and accidentals deteriorate the final image quality.
For a total-body scanner, the background level becomes a challenge.
First, the accidentals statistics increase roughly quadratic with the
scanner axial length. Second, the multiply scattered photons fraction is
more pronounced. Morover in J-PET scanner the signal registration is based on
the Compton scattering process, which makes the inter-detector scatters
harder to discriminate.
We apply supervised learning models to estimate the background
contribution. In particular, boosted decision trees and deep learning
neural networks are considered. The training and test samples are based
on GATE Monte Carlo simulations. Selection of optimal feature set and feature
transformations is performed. Performances of XGBoost, AdaBoost and
selected NN classifiers are compared with cut-based selection criteria.
Considered models are compared based on efficiency metrics. Finally,
preliminary comparison of reconstructed image quality is provided.
One of the biggest challenges in the deep learning application to the
medical imaging domain is the availability of training data. A promising
avenue to mitigate this problem is the usage of Generative Adversarial
Networks (GAN) to generate images to increase the size of training data
sets. A GAN is a class of unsupervised learning methods in which two
networks (generator and discriminator) are joined by a feedback loop to
compete with each other. In this process the generator gradually learns
how to better deceive the discriminator, on the other hand, the
discriminator gets constantly better at detecting synthetic images.
We will present the results of the transfer learning-based
classification of COVID-19 chest X-ray images. The performance of
several deep convolutional neural network models is compared. Data
augmentation is a typical methodology used in machine learning when
confronted with limited data set. We study the impact on the detection
performance of classical image augmentations i.e. rotations, cropping,
and brightness changes. Furthermore, we compare classical image
augmentation with GAN-based augmentation. A StyleGAN2-ADA model of
Generative Adversarial Networks is trained on the limited COVID-19 chest
X-ray image set.
After assessing the quality of generated images they are used to
increase the training data set, and to improve the balance between classes.
Radiotherapy aims at treating patients with cancer using ionising radiation. However, a key step is the optimization of the treatment. This is done using an inverse-planning approach where the treatment goals are encoded into a cost-function to minimize. The latter can be either non-convex or non-smooth with several local minima.
Quantum computers may efficiently solve this problem thanks to their inborn parallelisation ability. Therefore, in the last two years, our group focused on the development of new optimization strategies based mainly on Tensor Network Methods where
the classical optimization problem is mapped into an ising-type Hamiltonian whose ground state corresponds to the best solution to the initial problem and the optimization variables are represented in terms of qubits.
Our preliminary results show that this approach is compatible with any type of function and can perform at least comparably as classical optimization algorithms on the test functions considered.
Quantum algorithms are typically expressed as a quantum logic circuits, where the qubits of a quantum computer are sequentially operated by quantum gates. These gates are quantum mechanical counterparts of classical logic gates such as NOT, XOR etc. that enable more powerful computing by using quantum superposition and entanglement.
This hands-on training covers the following
Prerequisites
For this hands-on you will need to preinstall Python with packages qiskit and Jupyter notebooks for viewing and running the notebooks of the session. Instructions for installing qiskit can be found here: https://qiskit.org/documentation/getting_started.html and as a video here: https://www.youtube.com/watch?v=M4EkW4VwhcI (only first 4.5 minutes are relevant).
Installing conda (as instructed in the above links) is not compulsory, but might help things run smoothly. This hands-on serves as an introduction to quantum computing and does not require previous experience.
Quantum algorithms are typically expressed as a quantum logic circuits, where the qubits of a quantum computer are sequentially operated by quantum gates. These gates are quantum mechanical counterparts of classical logic gates such as NOT, XOR etc. that enable more powerful computing by using quantum superposition and entanglement.
This hands-on training covers the following
Prerequisites
For this hands-on you will need to preinstall Python with packages qiskit and Jupyter notebooks for viewing and running the notebooks of the session. Instructions for installing qiskit can be found here: https://qiskit.org/documentation/getting_started.html and as a video here: https://www.youtube.com/watch?v=M4EkW4VwhcI (only first 4.5 minutes are relevant).
Installing conda (as instructed in the above links) is not compulsory, but might help things run smoothly. This hands-on serves as an introduction to quantum computing and does not require previous experience.
Image reconstruction for positron emission tomography (PET) has been developed over many decades, starting out with filtered backprojection methods, with advances coming from improved modelling of the data statistics and improved modelling of the overall physics of the data acquisition / imaging process. However, high noise and limited spatial resolution have remained major issues in PET, and conventional state-of-the-art methods have exploited other medical imaging modalities (such as MRI) in order to assist in denoising and enhancing the spatial resolution for PET. Nonetheless, there is a drive towards not only improving image quality, but also to reducing the injected radiation dose and reducing scanning times. While the arrival of new PET scanners, such as total body PET (TB PET), is helping, there is still a need to improve the reconstruction of PET images in terms of quality and speed.
Deep learning methods are forming the new frontier of research for PET image reconstruction. They can learn the imaging physics and its inverse, learn the noise and also exploit databases of high-quality reference examples, to provide improvements in image quality. There are four main approaches: direct full data-driven learning of reconstruction operators, direct methods which incorporate known imaging physics, methods which integrate deep learning into existing iterative reconstruction algorithms (unrolled reconstruction) and methods which exploit deep learning as a means of representing the images to reconstruct (e.g. the deep image prior). This talk will cover a review of these methods, their advantages and disadvantages. The outlook of current and future directions for deep learning in PET reconstruction will then be considered, such as self-supervision, and quantifying uncertainty.
Convolutional Neural Networks (CNNs) have been effectively applied in many studies where crucial information about the data is embedded in the order of features (e.g. images). However, most tabular data – such as raw Positron Emission Tomography (PET) data – do not assume a spatial correlation between features, and hence are unsuitable for CNNs classification. In order to use the power of CNNs (including GPU utilization) for classification purposes of non-image data, a transformation method of 1-D vector into image has to be applied. A method comparison of transforming tabular data into input images for CNN classification will be presented. Self-organizing map and DeepInsight method were used in this study.
Big variety of medical data types and their complex structure may be a challenge for data scientists. The process of creating the data is usually time-consuming, while access to medical facilities databases is limited due to privacy issues.
Generative models can be of great help in the process of data augmentation. The presentation will contain the idea, current status and results of generative models training (AutoEncoders, GANs) in order to build a tool for generating medical data: 3D medical DICOM images representing the patient's geometry as well as phase space files necessary in the process of simulating the radiotherapy dose deposited in the phantom.
LUMI is the first pre-exascale supercomputer of the EuroHPC Joint Undertaking and is now Europe’s most powerful supercomputer. Finland, together with 9 other countries from the Nordics and central Europe, collaboratively hosts one of these systems in Kajaani, Finland. The vast consortium of countries with an established tradition in scientific computing and strong national computing centers is a key asset for the successful infrastructure.
The LUMI supercomputer is also one of the most advanced platforms for artificial intelligence. It links together computational capacity, artificial intelligence methods (especially deep learning), traditional wide-scale simulation and the utilization of large masses of data to simultaneously solve a single challenge. LUMI serves as well as a platform for the development of quantum technology. Quantum computers need supercomputers alongside them to harness their capacity to the right targets as a part of the research process. LUMI has so far been linked successfully with two quantum computers: the Swedish QAL 9000 and the Finnish Helmi.
In this talk we will discuss the LUMI infrastructure and its great value and potential for the research community.
Field Programmable Gate Arrays (FPGAs) offer unique features for High-Performance Computing such as natural parallelism, streamlined processing, and dynamic reconfiguration creating a relatively new concept of adaptive computing.
Modern device capabilities, high-level development techniques, and market adoption make them a powerful and interesting component for HPC hardware platforms.
In this talk, I will present a technology overview and current techniques for implementing Neural Networks on FPGA-based platforms.
The Polish National Competence Center for HPC built within the EuroCC project will be described during the talk. EuroCC is a pan-European project funded by the EuroHPC JU, a joint initiative between the EU, European countries and private partners to develop a World Class Supercomputing Ecosystem in Europe. Polish NCC is formed on the base of PLGrid Consortium members, which include all six Polish HPC centres: Cyfronet, Cyfronet, CI TASK, ICM UW, NCBJ, PSNC, WCNS. A short overview of the project and the current status will be provided, as well as a list of currently available services and competencies provided by the NCC partners.
The LUMI is one of the European pre-exascale HPC systems hosted by the LUMI consortium. The LUMI (Large Unified Modern Infrastructure) consortium countries are Finland, Belgium, Czech Republic, Denmark, Estonia, Iceland, Norway, Poland, Sweden, and Switzerland. This one-day tutorial presents a technical overview of the system's hardware configuration and programming level environment. The aim of the course is to popularize hardware design of the compute nodes and network and associated programming environment. This introductory material is meant to be a quick-start for those who consider access to the LUMI resources and brief introduction to the software tools available and capabilities of the hardware.
Prerequisites
An SSH client. Appropriate accounts will be created on 14th of September after participants registration.
The LUMI is one of the European pre-exascale HPC systems hosted by the LUMI consortium. The LUMI (Large Unified Modern Infrastructure) consortium countries are Finland, Belgium, Czech Republic, Denmark, Estonia, Iceland, Norway, Poland, Sweden, and Switzerland. This one-day tutorial presents a technical overview of the system's hardware configuration and programming level environment. The aim of the course is to popularize hardware design of the compute nodes and network and associated programming environment. This introductory material is meant to be a quick-start for those who consider access to the LUMI resources and brief introduction to the software tools available and capabilities of the hardware.
Prerequisites
An SSH client. Appropriate accounts will be created on 14th of September after participants registration.
The conference dinner will take place on the 15th of September in Restauracja Stolica, Szeroki Dunaj Street 1/3
Stare Miasto (Old Town).
http://tiny.cc/uy8zuz
In the era of astronomical "big data", with the amount of observational
data ever-increasing and about to increase by orders of magnitudes
during the next decade, machine learning has become not only a commodity
but also a necessity. At the same time, the application of machine
learning methods to astrophysical problems yields many specific
challenges. One of them is related to the fact that while the data to
which we want to apply these methods are often big, the available
training samples are small. Moreover, they are often not really
representative, in a way that may be difficult to quantify, which faces
us with a variety of extrapolation problems. More challenges are related
to the interpretability of the results, given the limited information we
can access. I will try to discuss the aims, difficulties and attempts to
overcome them, making use, among other things, of examples from the
research made in our extragalactic astrophysics group in NCBJ and UJ.
The upcoming large-scale surveys like LSST are expected to find approximately $10^5$ strong gravitational lenses by analysing data of many orders of magnitude larger than those in contemporary astronomical surveys. In this scenario, non-automated techniques will be highly challenging and time-consuming, even if they are possible at all. We propose a new automated architecture based on the principle of self-attention to find strong gravitational lenses and its advantages over convolution neural networks are investigated. From our study, we showed that self-attention-based models have clear advantages compared to simpler CNNs. They have highly competing performance in comparison to the current state-of-art CNN models. Moreover, introducing the encoder layers can also tackle the over-fitting problem present in the CNNs by acting as effective filters. In addition, we have also identified some new strong lens candidates from the Kilo Degree Survey (KiDS) using this new architecture.
The LHCb is one of the four main experiments discovering cutting-edge physics at the Large Hadron Collider. The effects of the harsh conditions of constant radiation require maximum thought and care. This includes research for new applications and novel algorithms that will help to understand and predict the behaviour of the vertex locator detector at LHCb. Our studies include methods based on calibration data from 2012-2018 and insights for the new (pixel-based) Velo detector in the upcoming data-taking runs at LHC. Those methods will be introduced to the detector monitoring software ecosystem.
Despite the large amount of data generated by the Large Hadron Collider
(LHC)
so far, searches for new physics have not yet provided any clear evidence of beyond the Standard Model (BSM) physics. Most of these experimental searches focus on exclusive channels, looking for excesses in specific final states.
However, new physics could manifest as a dispersed signal over many channels.
It therefore becomes increasingly relevant to attempt a more global approach to finding out where BSM physics may hide. To this end, we developed a novel statistical learning algorithm that is capable of identifying potential dispersed signals in the slew of published LHC analyses. Aiming to minimize theoretical bias, our approach is not constrained to a specific BSM scenario.
Instead, the algorithm is tasked with building candidate "proto-models", precursor theories to the Next Standard Model (NSM), from small excesses in the data, while at the same time remaining consistent with negative results on new physics.
In this talk, we explain the concept as well as technical details of the statistical learning procedure. We also present proof of concept results obtained when running the algorithm over our database that contains the results of 100 searches conducted at the LHC. Finally, we sketch out our vision of how the NSM could then be constructed from such protomodels.
Run 3 of the Large Hadron Collider (LHC) of the data-taking period poses unprecedented challenges to the computing models used in the high-energy physics experiments of the LHC accelerator. Only in the LHCb experiment, the luminosity has increased by a factor of five. Recent results show that deep learning solutions techniques can significantly improve the performance of the cluster reconstruction in calorimeters when high occupancy is expected. In this talk, we will review selected results of the LHC experiments and, in particular, focus on the investigated convolutional (CNN) and graph neural network (GNN) solutions for planar, LHCb-inspired calorimeters with hybrid granularities.
Currently, over 50% of the computing power at CERN’s GRID (>500 000 CPUs in 170 centres) is used to run High Energy Physics simulations. The recent updates at the Large Hadron Collider (LHC) create the need for developing more efficient simulation methods. In particular, there exist a demand for a fast simulation of the neutron Zero Degree Calorimeter, where existing Monte Carlo-based methods impose a significant computational burden.
We propose an alternative approach to the problem that leverages machine learning. Our solution utilises neural network classifiers and generative models to directly simulate the response of the calorimeter. In particular, we examine the performance of variational autoencoders and generative adversarial networks, expanding the GAN architecture by an additional regularisation network and a simple, yet effective postprocessing step.
Our approach increases the simulation speed by 2 orders of magnitude while maintaining the high fidelity of the simulation.
Transfer learning is a machine learning (ML) technique of reusing models with pre-trained knowledge obtained for a general ML task, and applying it to another, more specific ML task, with limited training data or computational resources. This hands-on training will cover the following topics in computer-vision-related problems:
Prerequisites
For this hands-on you will need a google account to access Google Colab service. Due to the time constraints only the simplest networks will be trained during the tutorial, larger models will be left to experiment with for the participants as a home assignemnt.
Transfer learning is a machine learning (ML) technique of reusing models with pre-trained knowledge obtained for a general ML task, and applying it to another, more specific ML task, with limited training data or computational resources. This hands-on training will cover the following topics in computer-vision-related problems:
Prerequisites
For this hands-on you will need a google account to access Google Colab service. Due to the time constraints only the simplest networks will be trained during the tutorial, larger models will be left to experiment with for the participants as a home assignemnt.