AI in Action
This article was originally published in Issue One of Discover: Cancer Research In Manchester. All articles are available to read on the MCRC website and a PDF version can be accessed through the links at the end of the page.
What is AI?
Artificial intelligence, or AI, is a type of technology that simulates human intelligence. It allows machines, particularly computers, to perform complex tasks that ordinarily require human thought and decision making.
To accomplish specific tasks, AI relies on learning from large amounts of data, guided by algorithms-a set of rules that help to direct the AI’s decision making. By analysing data freely, AI can uncover patterns and relationships, generating results and predictions on its own, without human input.
You may already be familiar with large language models like OpenAI’s ChatGPT or Google AI’s Gemini that can engage in human-like conversation. These models are trained on text data from digitised books, scientific research, legal texts and even social media to understand, generate and analyse human language to create a natural computer-human interaction.
But AI is not limited to text generation- there are other types of AI that are designed to tackle different tasks.
For example:
- Generative AI- a type of AI that can create new content such as text, images, videos and even music. Large language models fall into this category.
- Predictive AI- that uses patterns it has learnt from existing data to make predictions about future trends or patterns in new data.
AI is used in many different industries to automate and streamline tasks. For example, in finance, AI uses pattern recognition to detect fraudulent transactions. In manufacturing, AI is used to forecast demand and in the entertainment industry, streaming platforms use AI to suggest new content for users.
In scientific research, AI is being used to interrogate large datasets, generate new research questions, identify new drug targets and interpret medical images.
A role for AI in predicting breast cancer risk
PhD student Stepan Romanov uses predictive AI in his work aimed at improving risk prediction for breast cancer. His focus is on breast tissue density- a measure of the amount of fatty versus glandular and fibrous tissue in the breasts.
“My project aims to better predict a women’s risk of developing breast cancer by using AI to make the most of data available from mammograms from women enrolled in the NHS Breast Screening Program. Our goal is to make risk prediction more personalised, not just to detect breast cancer early, but also to predict the development of a cancer before it is even formed. Personalised risk prediction would also help us to identify women who are at low-risk who could be moved into the lower-risk program so that they are not being screened unnecessarily” Stepan explains.
The NHS Breast Screening Program.
Anyone registered with a GP as female will be automatically invited for an NHS breast screening, beginning between the ages of 50 and 53. Screening then takes place every three years, up until a person’s 71st birthday.
The screening uses a type of X-ray called a mammogram to check for breast cancer and can find a cancer even before there are visible signs or symptoms. Four mammograms are taken in total to get the best images of both breasts, taking a few minutes each. Results of the mammograms are sent through the post, usually within two weeks of the screening appointment.
In 2022-23, the screening program detected cancers in 18,942 women in England, and it is estimated to save 1,300 lives each year.
Around 56,400 women and 390 men are diagnosed with breast cancer every year in the UK.
How AI sees what we can’t
Typically, mammograms are used to detect lesions, abnormal growths within breast tissue and signs of breast cancer, and their role ends there. However, there is a wealth of information in these scans that could be useful for predicting cancer risk.
“It would be incredibly difficult for a radiologist to look through thousands of mammograms to see what features might correlate with a person’s risk, but this is where AI is really good” Stepan says.
Stepan’s work started by processing the mammograms into a usable format so that they could be applied to what’s called a convolutional neural network or a CNN, a specialised AI tool that works specifically on images.
Explainer: What is a CNN and how does it work?
A CNN is a mathematical tool that looks for broad features in an image, called the input.
Imagine you have an image of a cat. A CNN uses a filter to scan over the image in small sections like using a magnifying glass to look at a photo. Each filter is designed to extract features from the input image, like simple shapes, patterns and edges. Or, in the case of the cat, features like eyes, fur or whiskers. These are made into separate images showing the fur in one, the eyes in another etc.
As the CNN identifies more and more of these features, they are combined, and complex patterns and shapes begin to emerge, like the ears or face of the cat for example.
After processing the input image, the CNN connects all the identified features together and makes a prediction.
In the case of the cat, the CNN combines all the features it recognised, like the ears and the fur and it concludes that the image is of a cat.
How the CNN is used in research
The CNN identifies features, shapes and patterns within the mammograms and based on this, the AI returns a personalised risk score.
Accurate risk predictions rely on the model being thoroughly trained on a large and diverse data set. “In our study, we had a training set of mammograms from about 30,000 consenting women. We are really fortunate as we have a huge amount of data to work with just because the breast screening program exists!” Stepan says.
“To train the model, we take an image knowing the outcome of the scan, feed it through the model and compare its prediction to the known outcome. If it got it wrong, it tunes its parameters to predict the correct outcome. If you do this enough times, it will be able to correctly classify an unseen image based on what it has been previously taught”.
“In our work, we’ve seen that AI performs on par with radiologists at predicting risk from mammographic density.”
But don’t worry, the aim is not to replace your doctor or radiologist with AI, its role currently is to assist them. “AI could be used to provide a second opinion” Stepan explains. “It could look at a scan and identify features that a radiologist may have missed, and a benefit is that it is instant. After the model is trained, it is as simple as pressing a button and it would immediately tell you what the answer is”.
to assist them. “AI could be used to provide a second opinion” Stepan explains. “It could look at a scan and identify features that a radiologist may have missed, and a benefit is that it is instant. After the model is trained, it is as simple as pressing a button and it would immediately tell you what the answer is”.
The challenges
However, a major obstacle that stands in the way of using AI for risk prediction is what is known as the ‘black box’ problem. This is a fundamental problem with AI- that it works in ways that we do not fully understand, so we do not know how it works internally and how it reaches its conclusions, unlike a human that can explain their reasoning.
The problem with this, Stepan explains, is that “As AI makes the prediction based on your mammogram, we really can’t tell a patient anything more than their risk score and this could understandably frustrate patients. Whereas current risk prediction models that consider your basic clinical features can tell you that you are high risk because you have a gene alteration, or you are using hormone replacement therapy for example”.
Looking forward
AI has the potential to transform cancer care by offering a more informed, personalised approach to risk prediction. By analysing existing medical data, AI can help to identify patients at risk of developing cancers earlier, leading to earlier detection and patients being started on lifesaving treatments sooner, ultimately improving patient outcomes. Research like Stepan’s is making important strides towards integrating AI into everyday healthcare.
To learn more about the topics covered in this article, visit the National Cancer Institute, CoppaFeel, and NHS websites
