Creating fake lesions on brain MRI to train machine learning

May 31st, 2021


Brain microbleeds impact cognition and are associated with cerebrovascular disorders. Depending on location and number, they can lead to dementia and are often associated with cardio-vascular diseases.

Magnetic Resonance Imaging of the brain can reveal those microbleeds. They show as small dark spherical spots (a few mm in diameter) in the brain tissue when using specially designed MRI scans. Fortunately, they are quite rare: they are found in less than 10% of people in their 60s with a prevalence that increases with age.

Detecting those microbleeds from MRI is time consuming and tedious as specialist doctors need to visualize every one of the 150+ sections comprising a brain scan. Many normal features in the brain look like microbleeds such as blood vessels. Careful inspection by specialist of the 3D data is required to ascertain the presence of a lesion.

Machine learning has been investigated for automatically detecting those microbleeds with some successes. However, because of the variability in location, size, and appearance as well as the type of MRI, parameters used for scanning, etc.. it is very difficult to build a model that generalizes and can be used in different hospitals across many different patients.


Example of a brain microbleed (left) and a blood vessel (right) in a typical brain MRI


Our novel solution

We designed a model of microbleed that matches real lesions. This allowed us to create fake microbleeds of any size, location, appearance and add them to otherwise healthy MRI without any known lesions. We could thus create large datasets with many lesions, on many different patients from any MRI scanners. We demonstrated that training a machine learning method to detect microbleed using our synthetic data improve its performance when tested on real lesions.

Example of fake microbleeds of various sizes added at random on an MRI of a healthy subject



Learn more

This work is part of a large collaboration including the CSIRO Australian eHealth Research Centre, Griffiths University, and the Austin Hospital in Melbourne.

Our publication includes more details, and we created a dataset with 17,000 synthetic lesions available to researchers to train and benchmark novel AI methods.

CSIRO team: Saba Momeni, Amir Fazllolahi, Olivier Salvado.