Train schedules are a crucial part of our daily lives, but have you ever thought about the complexity behind creating them? From ensuring seats are available to deploying personnel; there are numerous factors to consider. Despite significant advancements in planning automation, there is still room for improvement, especially in hub planning optimization. So, what’s happening under the hood at NS, and why is optimizing railway planning so challenging? We interviewed Bob Huisman, Manager of Research and Development Hub Logistics at NS, to learn more about how AI is revolutionizing the train scheduling management process.

Bob Huisman is a respected figure in the railway industry, with his career spanning several decades. Huisman currently holds the position of Manager Research & Development Hub Logistics at NS, where he is responsible for delivering innovative methods and tools for planning and scheduling shunting related processes at railway hubs, as well as assessing the logistic process capacity of railway hubs. Beyond the technical and scientific aspects of his work, Huisman sees it as an opportunity to make a social contribution. He describes himself as ‘ having one foot firmly planted in the business world and the other in academia’. His  career path is a testament to his ability to bridge the gap between research, development, creativity, and challenging problems. Huisman is one the principal investigators of the LTP ROBUST programme and the chair of the user’s committee.

‘At first glance, the process of train schedule management may seem simple – travelers at the station, a train ready to board.’ But as Huisman points out, it’s actually a complex interplay of factors, while peering into the direction of Utrecht Central Station, one of the important hubs in the railway network. The hub planning, as part of the overall railway planning, involves ensuring trains are in the correct composition to maximize seat availability, that they arrive at the right platform at the right time, and that they are in good technical condition. Additionally, once a train reaches its final destination, it must be checked for technical issues, cleaned, potentially rearranged, and parked in a way that maximizes space efficiency.

But that’s not all. Railway planning and control involves fleet assignment and the deployment of personnel, which comes with its own set of complicated and important limiting factors. Employment conditions, work variation, and the ability of colleagues to come home at the end of the day are just some of the factors that the team of Huisman must take into account. ‘Train schedule management is almost paradoxical. As a traveler you may experience that we have one timetable during the entirety of the year, but under the hood, the rail sector makes a unique plan for every single day of the year. This involves planning for the timetable, fleet and staff, as well as for the 34 hubs – the stations with connected yards where multiple train lines converge – months in advance.’

Over the past two decades, NS has seen a significant increase in automation when it comes to network planning. However, there’s still no automation in hub planning, which Huisman notes as remaining an obstacle to overcome. Currently, this daunting task is solely on the shoulders of human hub planners, who are responsible for what is called the “knitting process”. ‘The “knitting process” involves juggling a multitude of factors simultaneously and making decisions in real-time, ensuring that passengers arrive at their destinations safely and smoothly.

How do those train scheduling experts manage to make everything run like clockwork?

‘It is a multi-step process. First, we create a timetable, then assign our fleet and lastly assign our colleagues.’ Albeit the largest, NS is only one of the 7 Dutch operators on the Dutch rail network, with ProRail responsible for infrastructure management, capacity allocation and traffic control. Huisman notes that planning NS’ train operation involves multiple iterations before arriving at the final basic pattern, which is then translated into a blueprint for each weekday, and subsequently each specific day. The schedule is finalized a month in advance for a specific calendar day, and from then NS still has the ability to make adjustments up to two days prior. ‘Once the plans are handed over for operational execution, controllers at ProRail and NS must act quickly in the event of a collision, malfunction, or employee illness. At that moment, it resembles tinkering more than actual planning by optimization’, Huisman notes.

NS has been making significant progress since 2015 towards automating the hub planning, but Huisman emphasizes that there are many “dirty details” that need to be taken into account. One of the most gratifying moments of this project came in 2020, just before the onset of the pandemic. ‘We were able to demonstrate on the basis of a proof of concept that we are going to make it’, Huisman recalls. ‘Moments like these; when colleagues have confidence in the success of a project and more resources become available; those keep me young. I have resolved that when I retire, to have this project standing,’ he says with a sense of pride. ‘This project represents a unique collaboration with young researchers, a continuous flow of PhD and master students, and has been one of the most personally rewarding projects of my career.’

Why is automation and optimization so important?

‘Two goals for the rail system are difficult to reconcile: on the one hand, meeting a growing demand for transport and, on the other hand, a robust operation. The driving force for automated hub-planning support is the need for fixing the plans as late as possible and be able to make changes on-line. That will improve the robustness of the transportation process, rail infrastructure usage, and seat availability,’ Huisman notes.

Currently, to anticipate uncertainties and unforeseen disruptions, slack is incorporated into the planning, with respect to both space and time. This slack allows the railroad planners to be flexible and to deal with details that will become clear on the day of operation itself; it gives space to breathe when things go wrong. ‘Reducing slack to facilitate future passenger volumes, increases the risk of a domino effect of disruptions, however, fast automated support for planning and control may compensate for this. All of these challenges beg the question; how do we achieve sustainable growth, act more dynamically and be more robust, while using our existing resources more efficiently?

NS seems to already have invested considerable resources into research that helps professional planners create more optimized plans in less time. Why is this process so tough?

‘Good question. Now, picture the entire rail system as a massive, interconnected wirework – a complex maze that requires meticulous planning to operate smoothly. There are countless variables that can impact the system, making optimization a daunting task. It’s not just a matter of flipping a few switches, pushing a few buttons and moving a few trains around’’.

‘Hub planning is a combinatorial problem with an enormous search space in which it is hard to find good and feasible solutions. Moreover, the need to fix plans as late as possible, requires to model many dirty details of the real world and complex safety rules, which excludes linear optimization methods. ‘When we started in 2015, hub planning had been the topic for a well-known international competition in the field of Operation Research. Curious as we were, we reached out to all the competition’s prize winners to see if they knew something we didn’t. Eventually we concluded that no practical solution was found yet, which was why we decided to set up a long-term research and development program ourselves. One way or another, we had to find a way. Together with our academic partners we finally succeeded in building a working system, mainly based on powerful local search, combined with other methods like linear optimization and constraint programming. We coined it the Hybrid Integrated Planning method (HIP). Although the systems generates plans that are acceptable to professional planners, continuation of the research is needed to enhance the system’s functionality.’

‘Following the progress made by Deepmind, preceding AlphaZero, we started a research track specifically focused on applying deep-reinforcement learning to multi-agent pathfinding. The idea was to build up a logistic plan by generating a chain of individual actions, representing train shunting movements. After six years of research together with our academic partners, we found that brute force local search still outperformed our various complex reinforcement learning approaches, even on simplified models of reality. Taking a step back to see the forest for the trees, we halted the program and shifted our focus. Our current research direction is aimed at using machine learning to complement local search, constrained programming, and linear optimization. The challenge is to find and modify plans more quickly, specifically plans that are easily understood by humans. To build a system that exhibits some intelligent behavior in the future, it must be able to learn from previous situations and to communicate at some abstract level with its users. We still have a long way to go, which asks for perseverance and creativity’, Huisman notes.

“Humans in the loop” does not conform to the traditional view on automation, in which AI systems entirely replace human operators. However, recently researchers have started to view automation as a two-dimensional process, where humans and machines work together and compliment each other’s strengths and weaknesses to achieve a common goal.

‘In our unique use case we are looking for automatic tools to support people’, Huisman emphasized. ‘Hub planning is a challenge that humans are quite good at and a task that requires substantial creativity, ingenuity and the ability to color outside the lines – but it takes a lot of time. On the other hand, algorithms can speed up the planning processes but currently cannot handle its full complexity. That is why we are focused on creating systems that help people to adjust the “knitting process” more efficiently; reducing slack and maximizing space usage to meet the growing demand for train travel.

When it comes to AI, Huisman has a unique perspective. ‘AI research can bring us closer to understanding human intelligence, yes. However, as long as we don’t understand and have defined human intelligence, stop talking about artificial intelligence that has to replace humans.’ Instead, Huisman believes that we should focus on building super tools that exhibit intelligent behavior, regardless of whether there’s a steam engine or neural network powering it. ‘I see neural networks and reinforcement learning as ingredients, among others, to create value’, Huisman explained, adding that it’s all about developing an overall system that can deal with the intricacies of logistic planning in cooperation with humans. ‘AI has the potential to disrupt the approach we take in this, but it is not just about generating plans automatically; you have to communicate the output to planners and make it understandable to them, give humans control over plan qualities, and link the output to the systems of other parties.’

Recently, ICAI announced the LTP ROBUST program; a new initiative supported by the University of Amsterdam and 51 partners from government, industry, and academia. As part of the program, 17 new AI labs are established that focus on the development of trustworthy AI technology to address socially relevant issues in areas such as healthcare, logistics, media, food, and energy. Can you elaborate on the importance of trust in AI systems and your role in the program?

‘Trust is a crucial factor in the adoption of AI systems. Our perspective is that the public, customers, travelers, patients, users and authorities all base their judgment not only on the functionality of the AI algorithm in isolation. Rather, they base it on the whole of interacting ICT, the organization behind it, the procedures put in place to regulate it, the UI, and the availability of the system. Trustworthiness of an individual AI algorithm is a necessary, but not sufficient condition for their effective use in a system. Therefore, research into creating such AI systems necessitates a symbiotic relationship between academia and industry. In the end, private or governmental organizations set the specifications of the system, design and build the system and operate the system over the years. In LTP, research, development and system engineering meet, to obtain social impact by operational trustworthy systems.

Trust is also contextual and domain-specific; the risks of a medical diagnosis differ from the risks of logistics planning or music recommendation, and people rely on different systems in different ways. The program’s approach is to start with a system vision and a targeted research question for each lab, with the private partners playing a vital role in validating the output and asking the right questions. ‘While the validation and questioning may vary for different fields, the general approach to winning the trust of the user can be similar. As one of the principal investigators and the chairman of the overarching user committee, my role is to oversee the cooperation between the partners and ensure knowledge transfer between labs.’

The RAIL Lab, a collaboration between Delft University of Technology, Utrecht University, ProRail and NS, is one of the LTP ROBUST Labs joining ICAI. Its goal? Working towards algorithmic support to ensure safe and reliable logistic operations and capacity planning that is trusted by human experts. Explainable AI plays a role in this.

‘Explainability is often seen from the research world as: if I could just explain why my algorithm came to this conclusion and if I change something about my input, how would my evaluation change? It’s almost an internal accountability from your algorithm to the outside world, which is necessary, but might only be sufficient to accept or reject an individual prediction of the system. The question is whether that is sufficient for humans to use and accept the system as a sparring partner for decision support.’

Huisman emphasized the importance of setting standards for what an explanation of an algorithm should look like. ‘Authorities often require a deeper understanding of how an algorithm works, including how it makes considerations, what information it looks at, what information is necessary to make a good choice, and how uncertain the algorithm is in its output. Furthermore, for specific instances, humans may ask counterfactual questions to understand why some decision is proposed and not some other. By understanding the requirements for human decision-making, we can create more effective explanations that provide a more complete understanding of the algorithm’s decision-making process. Since the user is often responsible for the final decision to be taken, he wants to be sure it is the right one.’

To address these challenges, each LTP ROBUST lab will include a researcher with a background in social and behavioral science. The RAIL Lab is a testament to this effort, with one PhD student focusing on the cooperation between human and AI planners. This study will reveal requirements, expectations, and potential pitfalls of human-AI interaction, specifically of interaction with algorithmic planners. These results will be augmented with data science techniques to extract important factors from past decision-making and planning processes, to develop a computational cognitive model of the decision and planning process.

Huisman sees a colorful future ahead: ‘NS has its fair share of critics – some say it’s too big, bureaucratic, or slow. On the other hand, I know fewer other companies that have invested as much time and resources into innovative projects like railway planning, as NS has in the Netherlands.’ Optimizing rail systems is a complex task that will require many more years of research and a delicate balance between human expertise and advanced AI algorithms. However, Huisman and his colleagues are committed and up for the challenge. ‘With LTP ROBUST and RAIL Lab’s ongoing efforts, we can hope to see more trustworthy, efficient and seamless rail systems in the near future.’

———

We hope that through this interview you learned a bit more about NS and the intricacies of railway planning. NS, ProRail and their academic partners, TU Delft and the University of Utrecht, are currently recruiting PhD students for the RAIL Lab. If you are interested in a complex technical AI challenge, in the light of a social contribution, check out their webpage: https://icai.ai/icai-labs/rail/ The next time you’re waiting for your train, take a moment to appreciate the intricate dance of 22.000 employees that’s happening behind the scenes to help you get to your destination, and maybe consider joining us!

The world is facing a number of converging climate change challenges: population growth, more frequent extreme weather events, and a need for the sustainable production of nutritious food. Some say that machine learning can support us to mitigate and prepare for such consequences of climate change, however, it is not a silver bullet. In this interview, Congcong Sun and Chiem van Straaten discuss the challenges of machine learning in agriculture and weather forecasting, and the similarities and differences between their respective fields.

On November 16th, 2022, ICAI organizes the ‘ICAI Day: Artificial Intelligence and Climate Change’ where Congcong, Chiem, and many other researchers will talk about how AI can be used to mitigate and prepare for the consequences of climate change. Want to join? Sign up!

	Congcong Sun  is an assistant professor in learning-based Control at Wageningen University & Research (WUR) and Lab Manager of the ICAI AI for Agro-Food Lab. Her research interests are in using learning-based control to explore the overlap between machine learning and automatic control and apply them to agricultural production.
	Chiem van Straaten is a PhD student at the Vrije Universiteit Amsterdam (VU) and the Royal Netherlands Meteorological Institute (KNMI). His research focuses on improving sub-seasonal probabilistic forecasts of European high-impact weather events using machine-learning techniques.

Congcong and Chiem, could you tell me what your research is about, and how it is connected to artificial intelligence?

Congcong: Yes, of course. My research focus is on learning-based autonomous control in agricultural production. For instance, in a greenhouse or vertical farm, climate control can be optimized to make the crops grow under more favorable conditions and produce a better quality crop. Another example is logistical planning for agro workers, such as harvesting robots in a multi-agent setting. Learning-based control applications are complex, which is why I mainly use deep reinforcement learning, which is the combination of reinforcement learning algorithms with neural networks.

Chiem: The research that I conduct pertains to studying and making predictions about weather and climate extremes. Many industries, such as agriculture production, depend on accurate weather forecasting. Understanding our climate better is crucial for preparing ourselves for extreme weather and at the same time allows industries to use their resources more efficiently. However, predicting weather events far in advance is extremely tough due to time lags, the conditional nature of observed patterns, and the multitude of factors influencing one another. Machine learning has the potential to deal with such levels of complexity, which is why I am interested in applying it to weather forecasting.

Do you see any similarities or differences between your research?

Congcong: I believe our research is interconnected. As Chiem mentioned, weather patterns are a large source of uncertainty within the agricultural industry, particularly for those applications where the farm is located in an uncontrolled environment, such as open-air farms.

In agriculture, the weather is not the only source of uncertainty, however. Uncertainty arises from the crops themselves. Different crops have optimal growing conditions, which means that a control policy that is effective for one crop might not be effective for another. Even if you were to place a different crop in the exact same greenhouse environment, you would need a vastly different policy for controlling it. What are your thoughts on that, Chiem?

Chiem: Yes, you are trying to tackle something that inherently is multivariate, which is similar to weather forecasting. Although I am not well-versed in the specifics of agriculture, I can imagine that you need to take into account many factors such as irrigation, lighting, and temperature?

Congcong: Yes, indeed. When we seek to regulate the climate within a greenhouse, there are a lot of variables we need to consider, like humidity, irrigation, fertilization, light, and temperature. Analyzing the relationships between these variables requires knowledge from various disciplines such as plant physiology and biology. Additionally, certain relationships might not have been discovered yet, which adds to the complexity of balancing these variables. The combination of machine learning and automatic control can help us explore some of these relationships and translate them into knowledge about how to best regulate these environments.

Chiem: Ah, exactly. Here, I see a great similarity between autonomous control of agriculture environments and the prediction of weather patterns. For a long time, physical numerical prediction models have been developed in order to incorporate as many of the processes that are known to be important for weather prediction as possible. However, it is also known that these models are not perfect, as the weather is extremely complex. Therefore, we attempt to replace parts of the numerical models with statistical models to capture yet-to-be-discovered processes

Congcong: Yes, indeed. What kind of data do you use to make weather forecasts?

Chiem: In the non-statistical forecasting models specifically, we use a plethora of data to make weather forecasts, including humidity, pressure, air temperature, and wind speed. Like the input, the output is often multivariate, similar to learning-based agriculture control. Another similarity might be that in both domains you encounter challenges due to cycles. For instance, I could imagine that in agriculture you need to take the growing cycle of plants into account, which is different for every plant. In weather forecasting, you also have to deal with many different cycles at the same time, such as the seasonal cycle, weekly cycles, and daily cycles.

Congcong: Yes, exactly! Plants have different optimal growing cycles. In greenhouses with multiple plants, it could be that different growing cycles overlap similarly to how cycles overlap in weather forecasting. It is interesting to see so many similarities between our two domains!

In your conversation, you mentioned some applications of machine learning in your respective domains. One challenge we often hear about is related to trustworthiness, especially in applications with high degrees of uncertainty. Are companies in your industry enthusiastic or reluctant to work with machine learning?

Congcong: Greenhouse climate control is quite mature in the Netherlands. Some commercial greenhouses have already implemented automated control, however, we are still not making use of all available cutting-edge sensing techniques. The adoption of such techniques by farmers might be slow since they are expensive and if they do not work as intended, it could ruin a farmer’s business. Also, farmers might be hesitant to trust machine learning technology, since it is a relatively new technology.

Chiem: As Congcong noted; the trustworthiness of a system is crucial for its widespread acceptance. Applications such as heatwave prediction are not quite ready for widespread use because heat waves have to be predicted far in advance, which is immensely tough to do accurately. Short-term forecasting applications, such as rainfall forecasting applications have a track record of successful predictions, however. Moreover, weather forecasting has rapid update cycles, so if you make an errant forecast today, you still have a chance tomorrow to forecast the same thing, but do so with greater accuracy. For heatwave prediction, such errant predictions have way more severe consequences. In agriculture, I could imagine the consequences are similarly more severe. What do you think Congcong?

Congcong: I agree with you Chiem. Plants are quite sensitive, so if a wrong prediction leads to hazardous conditions in which the plants cannot survive for long, the grower might lose all of their plants. While system control in agriculture does not come with direct harm to humans, like in autonomous driving, the margins on crops are small. Therefore, they are in general more averse to using machine learning and statistical modeling approaches in general.

Chiem, during the ICAI Day, a day revolving around the numerous challenges regarding machine learning and climate change, you will walk us through a heat wave prediction use case. What would you say the largest hurdle is in this research?

Chiem: The primary challenge in climate change research is the interaction between processes across different scales. On a local scale, processes such as heat exacerbation due to dry soil conditions or particular local atmospheric configurations can influence heat waves. However, such local conditions can also be synchronized across the scale of the complete northern hemisphere, which means that hundreds of kilometers away, very specific conditions might also be an indication of an impending heatwave. This can become increasingly more complex when you, for instance, include global connections.

The interaction across these many scales creates challenges in determining the resolution of the data you need and also what algorithm is most suitable to use. Additionally, climate change is actively changing our data distributions as we speak. Data that we gathered in the past might therefore have different weather dynamics than the weather right now, which makes generalizing very difficult. To an extent, your machine learning model is always extrapolating.

That is intriguing, thank you for your explanation! Congcong, during the ICAI Day you will moderate a lunch table discussion on Artificial Intelligence and Agriculture. What do you plan to discuss and why should people join?

Congcong: During the lunch table discussion, I would like to come together and talk about the current challenges of applying AI to agriculture; the popular and potential AI solutions to confront these challenges; as well as the future trends of applying AI to Agriculture. I believe it is valuable to join since it will be a very good chance for researchers, engineers, and students who are working in this area, or even just feel interested in this area, to ask their questions, share their opinions, and also may get some answers about their doubts through the discussions! Beyond that, it will also be a very good chance to build your network and explore potential collaborations for the future.

To round off; when would you say your research is a success?

Congcong: Any progress of my research, I consider a success and will make me happy. These are things such as my PhD students achieving a small step, solving pressing challenges for farmers, and making food production more sustainable by reducing emissions and energy use.

Chiem: One large success would be the ability to answer questions regarding climate change attribution such as: how much has climate change exacerbated the impact of this specific extreme weather event or made it more frequent? Being able to answer such questions confidently would allow us to hold parties, such as big emitters, accountable. While far off, I believe that machine learning has the potential to give us the tools necessary to do this in the future.

On November 16th, 2022, ICAI organizes the ‘ICAI Day: Artificial Intelligence and Climate Change’ where Congcong, Chiem, and many other researchers will talk about how AI can be used to mitigate and prepare for the consequences of climate change. Want to join? Sign up!

Medical imaging is a cornerstone of medicine for the diagnosis of disease, treatment selection, and quantification of treatment effects. Now, with the help of deep learning, researchers and engineers strive to enable the widespread use of low-cost medical imaging devices that automatically interpret medical images. This allows low and middle-income countries to meet their clinical demand and radiologists to reduce diagnostic time. In this interview, Phillip Lippe, a PhD student at the QUVA Lab, interviewed Keelin Murphy, a researcher at the Thira Lab, to learn more about the lab’s research and the developments of the BabyChecker project.

Keelin Murphy is an Assistant Professor at the Diagnostic Image Analysis Group in Radboud University Medical Center. Her research interests are in AI for low-cost imaging modalities with a focus on applications for low and middle-income countries. This includes chest X-ray applications for the detection of tuberculosis and other abnormalities, as well as ultrasound AI for applications including prenatal screening.

Phillip Lippe is a PhD student in the QUVA Lab at the University of Amsterdam and part of the ELLIS PhD program in cooperation with Qualcomm. His research focuses on the intersection of causality and machine learning, particularly causal representation learning and temporal data. Before starting his PhD, he completed his Master’s degree in Artificial Intelligence at the University of Amsterdam.

The QUVA Lab is a collaboration between Qualcomm and the University of Amsterdam. The mission of QUVA lab is to perform world-class research on deep vision. Such vision strives to automatically interpret with the aid of deep learning what happens where, when, and why in images and video.

The Thira Lab is a collaboration between Thirona, Delft Imaging, and Radboud UMC. The mission of the lab is to perform world-class research to strengthen healthcare with innovative imaging solutions. Research projects in the lab focus on the recognition, detection, and quantification of objects and structures in images, with an initial focus on applications in the area of chest CT, radiography, and retinal imaging.

In this interview, both Labs come together to discuss the challenges in deep learning regarding the medical imaging domain.

Phillip: Keelin, you witnessed the transition from simple AI to deep learning. What do you think deep learning has to offer in medical image analysis?

I believe deep learning has a huge role to play in medical image analysis. Firstly, radiology equipment is expensive and requires the training of dedicated physicians, which means that low and middle-income countries can not meet their clinical radiology demands. Using deep learning-powered image analysis, therefore, has the potential to homogenize medical imaging accessibility around the world.

Secondly, even in richer countries such as the Netherlands, we can use deep learning to reduce the costs of radiology clinics. Every minute a radiologist spends looking at an x-ray, for example, is expensive and radiologists have to review a lot of x-rays every day. While every x-ray still requires the radiologists’ utmost attention, many of these x-rays actually show no signs of malicious threat. Deep learning could be used here to prioritize radiologists’ work list, by putting cases that seem normal at the bottom and cases that are deemed urgent at the top of the list. When artificial intelligence can really be relied upon, we could even start removing items from the radiologists’ workflow entirely.

Phillip: You mentioned that you use deep learning, which of course has many facets of neural networks, such as graph neural networks (GNNs) or transformers. Since you are working in imaging analysis, I assume you mostly work with computer vision models. Are you using convolutional neural networks (CNNs) for classification and segmentation or do you even go beyond that scope?

As you mentioned, we almost always use a CNN, where the type of CNN is dependent on the application. More than not, due to the confidential nature of medical data, the most important factor in determining which model to use is actually the amount of data that is available. Using too little data to train a model runs the risk of overfitting and introducing lots of uncertainty. Therefore, we have to try to use models cleverly to make sure to overcome such consequences, by using class balancing, data augmentation, or adjusting the network architecture. Other factors include the size of the model. For instance, to enable global use of the BabyChecker, it is important that the model can fit on a mobile phone device which sets requirements for the size of the network we can use.

Phillip: We know that deep learning models can create false predictions, so it might happen that the system indicates that a measurement looks all good, while that person actually needs to go to the hospital. How do you deal with this uncertainty and possible mistakes?

First, we should acknowledge that uncertainty is inevitable. Radiologists make mistakes, just like models can make mistakes. Only through strict quality control processes can we ensure that these models are reliable enough so that they do more good than harm. Especially in the medical field, this poses many challenges. For instance, on the technical side, we should figure out how to deal with domain shifts. On the legal side, we should determine who is responsible if the model makes a mistake and what legal actions can be taken. Those things are incredibly unclear at the moment.

Right now, I still see that artificial intelligence has a big role to play as a suggestive assistant to a radiologist if one is present, or as a screening tool when one isn’t. For instance, in Africa tuberculosis is very prevalent but most often there is no physician available. One of the products developed in our group and now scaled by Delft Imaging is able to detect tuberculosis-related abnormalities in inexpensive chest x-rays and refer patients for a more accurate and expensive microbiological test when necessary. While this product is not flawless, it does allow us to help more people that we couldn’t have helped otherwise. So until we reach the stage that systems are sufficiently quality controlled, using deep learning for screening and suggestions can be really useful.

Phillip: This sounds similar to challenges in autonomous driving, where it is hard to determine who is really at fault in an accident. We know that another problem is that neural networks tend to be overconfident, also in situations where they should not be. Are there ways to address this problem?

Yes, I have not yet mentioned it to you before, but this is actually really important for getting artificial intelligence accepted in the clinical workflow. Sometimes it might happen that an image of noise accidentally makes its way into the database due to a malfunctioning scanner. If the system would still give you a score for emphysema, then you lose faith in that system. In such cases we want the system to output that the image is very different than the images it was trained on and that the model cannot classify that image. It would be even better if the system provides an interpretable explanation for why it made a certain prediction since transparency in the prediction process is crucial for clinicians to be able to trust the system.

Phillip: You mentioned interpretability, a topic that has gotten a lot of popularity recently. Especially due to discussions about whether interpretability techniques are truly interpretable. Have you already tried out interpretability methods for neural networks or are those methods still a bit too noisy?

While interpretability methods work well in theory, for me the field is still too under-researched to have practical value. One popular method for explaining predictions in the medical field is by producing heat maps based on the weights of the network. However, such methods are hard to quantify and look more pretty than actually being useful explanations.

Phillip: In the low data regime, where models are trained on small amounts of data, the explanations might also quickly overfit to random noise. Yes, indeed. When clinicians hear about these topics in AI, are they reluctant to participate in research on artificial intelligence?

My experience is really positive, but I work mainly with doctors who are interested in what artificial intelligence has to offer. I believe that clinicians recognize that AI is coming to their field and they either have to get on board or they are going to be left behind. As I mentioned before though, the technology in most cases is not ready to be left unattended. Therefore, use cases that researchers and clinicians prefer right now, often assume a suggestive/assistant role for the AI algorithm or a screening role in scenarios where no trained reader is available.

Phillip: Since we talked a lot about data sparsity, how important is it for you to have collaborations across hospitals or medical companies to get access to data?

Collaboration with your partners is super important, in lots of ways. I believe that researchers should never try to develop medical image analysis solutions if they are not collaborating with clinicians. First off, to do research, you are dependent on the availability and the quality of data that is gathered. If we want to move the field forward, we should communicate about the following topics more: which data can be used for what research purpose; how should the data be gathered so that the quality is the highest; and how can we get consent from patients to use their data more often than we do now.

Secondly, there is some knowledge sharing involved. Sometimes I read a paper that had no clinical input and you can really see the difference. Either the researchers made mistakes that could have been prevented, or you can see that the research does not have practical value.

Phillip: Do you consider the agreement of a patient to use their data to be the biggest hurdle for developing something like a medical ImageNet?

The problem is that patients are not asked often enough if their data can be used for commercial purposes. Even so, when they are asked, patients might be reluctant to share such private data without being aware of how and by who it is used. While everybody working in the field of artificial intelligence knows that data is the cornerstone of everything, we should think about how we can communicate this effectively to the community. For instance, by providing more education to create public awareness of what AI is and why large amounts of data are necessary to create successful solutions.

Phillip: From the perspective of a patient, it is a small thing to give, but for the research domain, every single patient who is willing to share their data makes a big difference in enabling better medical image analysis.

Yes indeed. Still, there are challenges that need to be addressed. For instance, do patients feel comfortable with sharing their data with all companies or do they prefer to share their data selectively? What does it mean for competition if all companies have access to the same data? These are questions that we need to find an answer to, together with the community.

Phillip: Yes maybe, patients even want to go as specific as to approve for which specific applications their data can be used for. As scientists, we of course assume that data will be used for good, but we need to make sure that data is really only used for beneficial applications and not applications that might harm people.

Yes, and we should also make sure that data is completely de-identifiable, so that the person the image is taken of, can never be traced back to that image.

Phillip: Now, what is the research focus of the Thira lab?

Our research focuses on two things: the scalability of existing methods in the medical domain, and the reliability of the new methods we are developing to make predictions. Whatever research we do, I would say the red line is always the clinical applicability of our solutions rather than developing pure theoretical knowledge.

Phillip: When developing a device like the BabyChecker, do you only use data acquired with that device to train the model, or is there some domain adaptation involved?

In general, in the minimum viable product stage, we only use data acquired with the actual device so no domain adaption is necessary. At this early stage, BabyChecker’s software works with a selected ultrasound probe so that early adopters in our projects can gain easy access to BabyChecker. Over 70 operators who have been trained to use BabyChecker are scanning pregnant women in Tanzania, Ghana, Sierra Leone, Ethiopia, and soon Uganda as well. The data comes back to our partner Delft Imaging, where experts keep a close check on how well the software is working and where physicians determine the quality of the data. This way we make sure that the system is rigorous and that patients get the correct care.

Phillip: You have already mentioned some future improvements to the BabyChecker, where do you want to be in four years?

At the moment, the BabyChecker checks a few things: 1) The gestational age of the baby to determine the estimated due date, 2) The position of the baby so that when the baby is in a breech position, the woman can make sure to deliver the baby in a hospital, and 3) The presence of twins since this is also a high-risk pregnancy where the woman should go to the hospital to deliver. Additionally, we are looking to perform placenta localization and detect the fetal heartbeat to discover possible pregnancy complications

Phillip: Let’s say that in four years the field of AI has made at least one or multiple steps forward. Where do you see that AI needs to improve, especially in the medical domain?

In general, I would like to see how we can use low-cost x-rays and ultrasounds for lots of other diagnoses. For example, heart failure or lung disease. However, in order for such applications to be feasible, we need AI methods that can work well with small amounts of training data. I think that is really the biggest challenge that we have to overcome.

Phillip: In terms of evaluation, when would you consider your research to be successful? Is it when doctors use the products that you have developed or is it when you feel like there is nothing to improve in the short term?

While I believe I will never feel like there is nothing to improve, I would say my research is successful if we can reliably screen large amounts of people in low-resource settings for all sorts of illnesses and possible complications and get them referred for the treatment they need.

On October the 6th, 2022, the Thira Lab and the QUVA Lab will talk about their current work during the lunch Meetup of ‘ICAI: The Labs’ on AI for Computer Vision in the Netherlands. Want to join? Sign up!