Training AI to detect brain tumours

AI can help human experts catch brain tumours faster. That’s a fact, and it’s going to save thousands of lives. In this piece, we’re going to showcase how we’re training AI to detect brain tumours and why this is just the beginning.

Why are brain tumours hard to diagnose?

Brain tumours are a nightmare to diagnose. They are abnormal sizes, in irregular locations and present differently, with often very subtle symptoms. That means, until now, the NHS and other services have needed specialists to do time-consuming manual analysis of MRI scans. That’s a real problem for developing or deprived places where specialists are scarce. Add to that, there’s always a risk of human error, and it’s doubly hard to catch early-stage abnormalities, which might not even be that noticeable to the human eye.

How can AI help diagnose brain tumours?

Our models are trained to look at images of brain MRI scans and find abnormalities, no matter how small. With machine learning and AI in medical imaging, healthcare centres can process hundreds more images per day, increasing the chance that problems will be caught sooner. The results? Faster analysis, better pattern recognition and a tech that’s scalable to collect data across regions and act as a central processing hub.

What is multi-label classification?

The technology takes in MRI image datasets covering all 3 tumour types and outputs more accurate results than manual classification. Here’s how it works: By using a labelled dataset with MRI image data from all 3 brain tumour types and loading the MRI image data into Agxio’s Apollo platform, it is possible to train multiple models in parallel with deep learning to classify the type of tumour. Our best-performing model got an accuracy of up to 83% with some tuning and a ROC AUC of 0.96. This result is in line with the leading specialist doctors’ accuracy. The models can then be exported from Apollo and placed into a production system.

Why does it matter?

Doing things faster and more accurately represents both cost and time efficiency. This is great for under-resourced healthcare systems like the NHS, which already struggle with staffing issues. While humans aren’t removed from the process (and we’d never condone that anyway), AI can speed up diagnosis and improve treatment outcomes, reducing the workload for specialists and allowing them to focus on where they’re needed most.

What’s the future for AI in medicine?

This is just the first level. There are further efficiencies and benefits to be gained with continuous improvement and retraining of models. Just like we’ve shown with skin cancer and tumours, there’s broader potential for AI and ML to help with diagnosing other types of cancer and diseases too.

If you’d like to see first-hand how we’re training AI to detect brain tumours and other types of cancers, please get in touch for a demo today.