Tag: data science

The Royal Statistical Society Annual Conference 2024

Posted on by kerry.turcsi

The Royal Statistical Society meets annually for their internationally attended conference. It serves as the UK’s annual showcase for statistics and data science. This year they met in Brighton for a conference attended by over 600 attendees from around the world, including JGI Data Scientist Dr Huw Day.

The conference had over 250 presentations, including contributed talks, rapid-fire talks, and poster presentations. At any one time, there could be as many as 6 different talks going on, so it was impossible to go to everything but below are some of Huw’s highlights of the conference.

Pre-empting misunderstandings is part of trustworthy communication

From left to right; Dr Huw Day, Professor Sir David Spiegelhalter and Dr Simon Day
From left to right; Dr Huw Day, Professor Sir David Spiegelhalter and Dr Simon Day (RSS Fellow and Huw’s dad) at the RSS International Conference 2024.

As part of a session on communicating data to the public, Professor Sir David Spiegelhalter talked about his experiences trying to pre-bunk misinformation when displaying data.

Data in June 2021 showed that the majority of COVID deaths are in the vaccinated group. The Brazilian president President Jair Bolsonaro used this data to support claims that Covid vaccines are killing people. Spiegelhalter and his colleague Anthony Masters tried explaining why this wasn’t a sign the vaccine was bad in an article in The Observer “Why most people who now die with Covid in England have had a vaccination”.

Consider the following analogy: most car passengers who die in car accidents are wearing seatbelts. Intuitively, we understand that just because these two variables are associated, it doesn’t mean that one causes the other. Having a story like that means you don’t have to talk about base rates, stratification or even start to use numbers in your explanations.

We should try to make the caveats clearer of data before we present them. We should be upfront from what you can and can’t conclude from the data.

Spiegelhalter pointed to an academic paper: “Transparent communication of evidence does not undermine public trust in evidence” where participants were shown either persuasive or balanced messages about the benefits of Covid vaccines and nuclear power. It’s perhaps not surprising to read that those who already had positive opinions about either topic continued to have positive views after reading either messages. Far more interesting is that the paper concluded that “balanced messages were consistently perceived as more trustworthy among those with negative or neutral prior beliefs about the message content.”

Whilst we should pre-empt misconceptions and caveats, being balanced and more measured might prove to be an antidote to those who are overly sceptical. Standard overly positive messaging is actively reducing trust in groups with more sceptical views.

Digital Twins of the Human Heart fueled Synthetic 3D Image Generation

Digital twins are a digital replica/simulator of something from the real world. Typically it includes some sort of virtual model which is informed by real world data.

Dr Dirk Husmeiser at the University of Glasgow has been exploring the application of digital twins of the human heart and other organs to investigate behaviour of the heart during heart attacks, as well as trying to use ultrasound to measure blood flow to estimate pulmonary blood pressure (blood pressure in the lungs). Usually, measuring pulmonary blood pressure is an extremely invasive procedure, so using ultrasound methods has clear utility.

One of the issues of building a digital twin is having data about what you’re looking at. In this case, the data looks like MRI scans of the human heart, taken at several “slices”. Because of limitations in existing data, Dr Vinny Davies and Dr Andrew Elliot, (both colleagues of Husmeiser at the University of Glasgow)have been attempting to develop methods of making synthetic 3D models of the human heart, based on their existing data. They broke the problem down into several steps, working to generate synthetic versions of the slices of the heart (which are 2D images) first.

The researchers were using a method called Generative Adversarial Networks (GANs), where two systems compete against each other. The generator system generates the synthetic model and the discriminator system tries to distinguish between real and synthetic images. You can read more about using GANs for synthetic data generation in a recent JGI blog about Chakaya Nyamvula’s JGI placement.

Slide on “Generating Deep Fake Left Ventricle Images for Improved Statistical Emulation”.
A slide from Dr Vinny Davies and Dr Andrew Elliot’s talk on “Generating Deep Fake Left Ventricle Images for Improved Statistical Emulation”. The slide depicts how progressive GANs work, where the generator learns how to generate smaller, less detailed images first and gradually improves until it can reproduce 2D slices of MRIs of the human heart.

Because the job of the generator is far harder than that of the discriminator (consider the task of reproducing a famous painting, versus spotting the difference between an original painting and a version drawn by an amateur), it’s important to find ways to make the generator’s job easier early on, and the discriminator’s job harder so that the two can improve together.

The researchers used a method called a Progressive GAN. Initially they gave the generator the task of drawing a lower resolution version of the image. This is easier and so the generator did easier. Once the generator could do this well, they then used the lower resolution versions as the new starting point and gradually improved the correlation. Consider trying to replicate a low resolution image – all you have to do is colour in a few squares in a convincing way. This then naturally makes the discriminator job’s harder, as it’s tasked with telling the difference between two, extremely low resolution images. This allows the two systems to gradually improve in proficiency.

The work is ongoing and the researchers at Glasgow are looking for a PhD student to get involved with the project!

Data Hazards

On the last day of the conference, Huw alongside Dr Nina Di Cara from the School of Psychology at the University of Bristol presented to participants about the Data Hazards project.

Participants (including Hadley Wickam, keynote speaker and author of the famous R package tidyverse) were introduced to the project, shown examples of how it has been used and then shown an example project where they were invited to take part in discussions about which different data hazards might apply and how you might go about mitigating for those hazards. They also discussed the importance of focussing on which hazards are most relevant and prominent.

Dr Huw Day (left) and Dr Nina Di Cara in front of a screen that says 'Data Hazards Workshop'
Dr Huw Day (left) and Dr Nina Di Cara (right) about to give their Data Hazards workshop talk at the RSS International Conference 2024.

All  the participants left with their own set of the Data Hazard labels and a new way to think about communicating hazards of data science projects, as well as invites to this term’s first session of Data Ethics Club.

Using Machine Learning to Correct Probe Skew in High-frequency Electrical Loss Measurements 

Posted on by kerry.turcsi

JGI Seed Corn Funding Project Blog 2023/24: Jun Wang & Song Liu

Introduction 

This project develops a machine learning approach to address the probe skew problem in high-frequency electrical loss measurements. 

A pipeline using ML model to correct the probe skew in measuring a magnetic hysterysis loop. Model training and model deploy method shown
Fig.1 (main figure) A pipeline using ML model to correct the probe skew in measuring a magnetic hysterysis loop 

What were the aims of the seed corn project? 

To tackle the net-zero challenge through electrification, power electronic converters play an important role in modern electrical systems, such as electric vehicles and utility grids. Accurate characterisation of individual component’s loss is essential for the virtual prototyping and digital twins of these converters. Making loss measurements requires using two different probes, one voltage and one current, each with its own propagation delay. The difference in the delays between the probes, known as skew, causes inaccurate timing measurements which leads to incorrect loss measurements. Incorrectly measured loss will misinform the design process and the digital twin, which can lead to wrongly sized cooling component and potential failure of the converter systems in safety-critical applications, e.g. electric passenger cars.  

As the aim of this project, we proposed to develop a Machine Learning based solution learn from experimentally measured datasets and subsequently generate a prediction model to compensate the probe skew problem. This interdisciplinary project treats the challenge as an image recognition problem with special shape constraints. The goal is a tool developed for the engineering community which takes in raw measurements and outputs the corrected data/image.  

What was achieved? 

Joint research efforts were made by the interdisciplinary team across two schools (EEME and Mathematics) for this project with the following achievements: 

  1. We have explored the options and made a design choice to utilize an open-source database as the foundation for this project (MagNet database produced by PowerLab Princeton), which provides rich datasets of experimentally measured waveforms. We then have developed an approach to artificially augment the data to create training data for our ML model. 
  1. We successfully developed a shape-aware ML algorithm based on the Convolutional Neural Network to capture the shape irregularity in measured waveforms and find its complex correlation to the probe skew in nanoseconds. 
  1. We subsequently develop a post-processing approach to retrospectively compensate the skew and reproduce the corrected image/data.  
  1. We evaluated the proposed ML-based method against testing datasets, which demonstrated a high accuracy and effectiveness. We also tested the model on our own testing rig in the laboratory as a real-life use case. 
  1. We have developed a web-based demonstrator to visualise the process and showcase the correction tool’s capability to the public. The web demonstrator is hosted on Streamlit and accessible through this link. 
Snapshot of the web application demo showing phase shift prediction at 620 test index
Fig.2 Snapshot of the web application demo 

Future plans for the project 

This completed project is revolutionary in terms of applying ML to correct the imperfection of hardware instruments through software-based post-processing, in contrast to conventional calibration approaches using physical tools. This pilot project will initiate a long-term stream of research area leveraging ML/AI to solve challenges in power electronics engineering. The proposed method can be packaged into a software tool as direct replacement/alternative to commercial calibration tools, which cost ~£1000 each unit. Our plans for the next steps include 

  1. Create documentations for the approach and the pipeline 
  1. Write a conference/journal paper for dissemination 
  1. Explore the commercialisation possibilities of the developed approach 
  1. Further improve the approach to make it more versatile for wider use cases 
  1. Evaluate the approach more comprehensively by testing it on extended sets of data 

Contact details and links 

Dr Jun Wang, Jun.Wang@bristol.ac.uk 

Dr Song Liu, Song.Liu@bristol.ac.uk 

Web demo: https://skewtest-jtjvvx7cyvduheqihdtjqv.streamlit.app/ 

The project was assisted by University of Bristol Research IT.

Understanding the growth of human brains 

Posted on by kerry.turcsi

JGI Seed Corn Funding Project Blog 2023/24: James Armstrong

The human brain is a highly complex structure and an inaccessible organ to study, which has hampered our understanding of how the brain grows, becomes diseased, and responds to drugs. In the last ten years, a new method has been developed that uses stem cells to grow miniature brain tissues in the lab. These “brain organoids” have proven to be an incredibly useful tool for scientists studying the human brain. 

However, a well-known limitation of this tissue model is their unpredictable growth: within the same batch, some organoids will undergo typical neural development with large cortical buds (Figure 1A) while others will fail to produce these important structural features (Figure 1B). This Jean Golding Institute funded project sought to answer the question – can do seemingly identical stem cell cultures undergo such different growth? To this end, we aimed to track the growth of ~600 brain organoids over 20 days, then to use computer vision / machine learning methods to pick out key structural features that could be used to predict the tissue growth. 

Two brain organoids grown using different methods and showing the growth from day 3 to day 20
Figure 1. (A-B) Examples of two brain organoids, grown using the same methods, that were identical at day 3 but undergo very different growth. (C) An example of the images acquired during this project. 

This work was led by Dr James Armstrong, a Senior Research Fellow who runs a tissue engineering research group at Bristol Medical School (www.TheArmstrongGroup.co.uk). Members of his team (Martha Lavelle, Dr Aya Elghajiji, with help from Carolina Gaudenzi) have so far grown ~200 organoids, with another member of his team (Sammy Shorthouse) collecting microscopy images at ten intervals throughout the growth (Figure 1C). As expected, we saw tremendous variation in the growth of the brain organoids, in terms of their size, shape, and budding. Sammy has developed a program that takes these images and automatically processes them (indexing, identifying correct focal plane, centring, and cropping). He is now developing this script into a user-friendly “app”. For the next stages, Dr Qiang Liu in the Faculty of Engineering has been working with Sammy to develop computer vision methods that can pick out the key structural features of the organoids at the different stages of their growth. We are now growing the next batch of organoids and hope to reach the ~600 mark by the end of the summer. This should provide us with our target dataset, which should be large enough to start drawing links and making predictions of tissue growth. 

Contact details and links

If you wish to contact us about this study, please email james.armstrong@bristol.ac.uk

Chakaya Nyamvula’s JGI Placement 

Posted on by kerry.turcsi

Hi, I’m Chakaya. I am currently pursuing my MSc in AI and Data Science at Keele University and working as a Business Intelligence Analyst at iLabAfrica at Strathmore University in Nairobi, Kenya. This summer, thanks to the partnership between iLabAfrica and JGI, I had an amazing opportunity to work with JGI for my Master’s placement. I wanted to immerse myself in a research environment and connect with people in academia to help figure out my future career path. Working under the guidance of Dr Huw Day, I gained valuable insights into the world of research and expanded my professional network, all while experiencing life in the UK. 

Chakaya Nyamvula in front of a body of water
Chakaya Nyamvula, JGI Intern

What was the project about? 

Previously for a JGI funded Seedcorn project Mark Mumme, Eleanor Walsh, Dan Smith, Huw Day, and Debbie Johnson had surveyed researchers on their thoughts on how they might want to use synthetic data to help with their research. 

Synthetic data is when you take an existing dataset and create a synthetic (i.e. fake) version of it. You might want to do this so you can share something that looks like the data but preserves the privacy of individuals in it, whilst still having a flavour of what the data looks like and what statistical patterns might be present within it. This is useful for writing data pipelines whilst you go through necessary ethics checks to access sensitive data, amongst other things. 

For my summer placement with JGI, I worked with the MIMIC IV dataset of electronic health records and explored methods of generating synthetic versions of some of this data. It was also important to understand how you could measure or benchmark how successful your synthetic data generation has been, based on how well you had preserved privacy or how well the statistics of your synthetic data emulated those of your real data. 

What else did you do as part of your placement? 

Alongside my main work, I attended JGI Data Science meetings and learnt about some of the data science projects at the JGI including a project on antimicrobial resistance and another on 3D image analysis of CT scanned zebrafish to study bone development. 

For some of the more computationally demanding aspects of the project, I got taught how to make use of the JGI’s server (known within the office as “Jeeves”). 

I also had the opportunity to meet some PhD students at the University of Bristol, ask them about their research, and get advice on applying for PhDs in the future. 

Left to right, Huw Day, Elena Fillola Mayoral, Yujie Dai and Chakaya Nyamvula sat at a table at an ice cream shop
From left to right: Huw Day (JGI Data Scientist), Elena Fillola Mayoral (PhD student in AI for Climate), Yujie Dai (CDT in Digital Health) and Chakaya Nyamvula (JGI Intern) discussing PhDs over ice cream

What did you learn about? 

One deep learning method we used was something called a Generative Adversarial Network (GAN). Prior to this project, I had never worked with GANs before, so diving into this methodology was both challenging and exciting.  

A GAN works by having two competing neural networks, a generator and a discriminator. The generator’s job in this case was to take the original data and generate synthetic versions of that data. The discriminator’s job is to try and spot the difference between the real and the synthetic data that has been generated. One of the advantages of such a system is that you have two outputs: 1) a neural network which can generate synthetic data based on some training data and 2) a second neural network which can discriminate between real and synthetic data. This has advantages for applications where people might maliciously generate synthetic data, for example deep fake images. 

A good analogy for GANs is two people learning chess by playing against one another. If both start at similar skill levels, then as one person improves, the other slowly improves too. If you lose a chess game, you know you made a mistake and you might be able to work out how to improve for the next time. If you win, then you know you were doing something right.  

However, if you pit a chess grandmaster against a complete beginner, then the beginner will lose every time and will struggle to understand where they are going wrong, making it difficult to improve. Because the task of making synthetic data is quite complicated, when we began the process of training the GAN, the generator was frequently getting it wrong and wasn’t really able to figure out how to improve. 

To combat this, we did two things. First, you can handicap the discriminator a bit to give the generator a head start (imagine making your grandmaster play blindfolded). This helped, but still wasn’t enough. 

One of the pair plots showing generated vs real data a epoch 0
One of the pair plots showing generated vs real data a epoch 25000
Pair plots showing how well the real and the synthetic data matches by comparing each column. Real data is in blue, synthetic data is in red. The diagonal plots show histogram density plots of each column and how it compares between real and synthetic data. The off diagonal show scatter plots between pairs of variables. The left pair plot shows the output at the start of training, where the synthetic generator just randomly samples a scatter of points. You can see that this is not a good match for the original data. The right pair plot shows that after training, the generator does a lot more of a convincing job at emulating the real data. It is still not perfect, but it is particularly good at identifying clumps of data.

Secondly, you can start to think about how you inform your neural networks whether or not they were successful. Imagine if instead of “win” or “lose” as your outcome of the chess games, you got a measure of how well you performed, say a measure of how many good moves you made. With this more specific information, it becomes easier to decipher why you lost and how you might improve.  

To Be Continued? 

To finish my placement, I shared my experience with my placement supervisors at Keele University through a presentation and a report. I then had the opportunity to present my work to the Data Science Seminar at the University of Bristol, with several lecturers from the data science community in attendance, alongside JGI Data Scientists and some friends I made along the way.  

Additionally, all the code we worked on can be found in a public GitHub repository for other researchers to use and experiment with can be found on Chakaya’s Github.

Chakaya Nyamvula and Huw Day standing in front of a projector presenting at the Data Science Seminar. The projector has a slide on it that says 'Introduction to synthetic data' 
Chakaya Nyamvula (left) and Huw Day (right) presenting at the Data Science Seminar 

Reflecting on my placement at JGI, I can confidently say it was an incredible learning experience. I had the privilege of working with a fantastic supervisor, Dr Huw Day, who provided guidance throughout the project. Co-working with the talented data scientists at JGI was both inspiring and rewarding, and I thoroughly enjoyed networking with professionals in academia. The challenges I faced particularly working with GANs for the first time, pushed me to grow and expand my skill set.  Overall, this experience not only deepened my technical expertise but also solidified my interest in pursuing a career that bridges research and data science. 

Working towards more universal skin cancer identification with AI 

Posted on by kerry.turcsi

JGI Seed Corn Funding Project Blog 2023/24: James Pope

9 examples of malignant/benign cancer marks on different skin types
Images from the International Skin Imaging Collaboration (https://www.isic-archive.com/

Introduction

Open-source skin cancer datasets contain predominantly lighter skin tones potentially leading to biased artificial intelligence (AI) models. This study aimed to analyse these datasets for skin tone bias. 

What were the aims of the seed corn project? 

The project’s aims were to perform an exploratory data analysis of open-source skin cancer datasets and evaluate potential skin tone bias resulting from the models developed with these datasets.  Assuming biases were found and time permitting, a secondary goal was to mitigate the bias using data pre-processing and modelling techniques. 

What was achieved? 

Dataset collection

The project focused on the International Skin Imaging Collaboration (https://www.isic-archive.com/) archive that contains over 20 datasets totalling over 100,000 images.  The analysis required that the images provide some indication of skin tone.  We found that only 3,623 recorded the Fitzpatrick Skin Type on a scale from 1 (lighter) to 6 (darker).  For each image, we mapped the Fitzpatrick Skin Type to light or dark skin tone.  As future work, the project began exploring tone classification techniques to expand the images considered. 

Artificial Intelligence Modelling

We then developed a typical artificial intelligence model, specifically a deep convolutional neural network, to classify whether the images are malignant (i.e. cancerous) or benign. The model was trained from 2/3 of the images and evaluated in the remaining 1/3.  Due to computational limits, the model was only trained for 50 epochs. The model’s accuracy (how many correct classifications it made of either benign or malignant tumours out of all the tumours it was evaluated on) was comparatively poor with only 82%. 

Bias Analysis

The model was then evaluated relative to light and dark skin tones.  We found that the model was better at identifying cancer in light versus dark skin tone images.  The recall/true positive rate for dark skin tones was 0.26 while for light skin tones it was 0.45.  The resulting disparate impact (a measure used to indicate if a test is biased for certain groups) was found to be 0.58, which indicates the model is potentially biased.

Future plans for the project 

The project results were limited due to the subset of images with skin tone and constrained computational resources.  Future work is to further develop the tone classifier to expand the number of labelled images. Converting colour values from images into values more closely related to skin tone and then comparing with the tone labels of the image, might help train an AI model to exclude the tumour itself when classifying skin tone of the whole image. This is important as we know that the tone of tumours themselves is often different to that of the surrounding skin.

Heat map showing where the skin tone matches the label
An example image from ISIC which had its Fitzpatrick Skin Type labelled. The light green indicates where individual pixels correspond with expected colours associated with the labelled skin type. Notice that the centre of the image, where the tumour is, does not match.

More powerful computational resources will be acquired and used to sufficiently train the model.   Future work will also employ explainable AI techniques to identify the source of the bias. 

Contact details and links 

James Pope: https://research-information.bris.ac.uk/en/persons/james-pope,

Ayush Joshi https://research-information.bris.ac.uk/en/persons/ayush-joshi,