The research team used a complex form of AI called deep convolutional neural networks (CNN) and trained it to recognise different types of brain tumours, as well as identifying specific features and patterns in images of brain tumours.
They exposed the AI programme to 838,644 images of brain tumours samples from 1,027 patients. Researchers added in associated notes about the patient’s diagnosis (by a pathologist), survival and therapy response.
The AI tool was then able to pool all this knowledge and learn how to differentiate between brain tumours and detect even the smallest of patterns.
Historically, a brain tumour diagnosis has been made solely by a pathologist looking down a microscope at a small tumour sample.
This visual diagnosis is subjective, so the diagnosis could vary depending on the person viewing the sample. With advances in technology, including molecular testing and AI, this is changing.
The goal of Dr Diamandis and his team is to ultimately create a computerised, cloud-based AI tool that will be available for any clinician to use. The tool will help decrease the time and any errors that are involved in making a diagnosis.
Conversely, experts have cited that CNN (the form of AI used in this study) has limited application in a healthcare setting, as it can only accomplish very selective and narrow tasks.
Additionally, they weren’t able to understand the rationale of how the images were grouped together and classified, as well as how a diagnosis was made.
Dr Diamandis and his team were able to overcome this limitation and prove the value of CNN by demonstrating that this form of AI was able to conduct more overarching and interpretive tasks.
The tool was able to group brain tumour images based on specific features, which are similar to features used by pathologists to make a diagnosis. Additionally, the tool was also able to find new patterns that aren’t recognisable by the human eye.
“Computers are inherently better at analysing large amounts of patient data. By being able to tune AI tools to focus on aspects we understand to be important to doctors, scientists will be able to discover relevant patterns in large image datasets in a synergistic way.
“By outlining the overlap between machine and human learning, we are optimistic that we are much closer to one day using these powerful computer tools in the care of patients with brain and other types of cancers,” says Dr Diamandis, lead researcher.
Tools such as CNN will help improve our understanding and diagnosis of brain tumours by automating the classification of brain tumour samples, reduce diagnosis error, and identify new patterns.