Breakthrough in Brain Imaging: New Method Creates Synthetic Scans to Aid in Diagnosing Brain Abnormalities
Diagnosing diseases like brain tumours often relies on medical imaging techniques like MRI scans. However, a new study by researchers at the School of Biomedical Engineering & Imaging Sciences has introduced a promising technique that creates synthetic brain MRI scans to improve the generalisability of medical segmentation algorithms.
This research, published in the Medical Image Analysis journal, could address the challenge of limited real-world data for training image segmentation algorithms, potentially leading to more accurate diagnoses and improved patient care.
Traditionally, clinicians rely on real patient data to train image segmentation algorithms, which are critical for identifying different brain structures and abnormalities. However, this approach is limited by the availability of real-world data, especially for rare medical conditions. This can lead to inaccurate diagnoses, particularly when dealing with less common diseases.
The new study proposes a generative model called brainSPADE that tackles this challenge by creating synthetic brain MRI images and labels. This essentially provides a way to train segmentation algorithms on a vast amount of artificially generated data. brainSPADE functions in two stages: first, it generates a map indicating various brain tissues and potential abnormalities. Then, it utilizes this map to create a realistic-looking MRI image.
In this work, we have shown that guided labelled synthetic data can indeed help downstream machine learning algorithms performing better in the presence of phenotypes that are not seen in the real training data. Although this paper addresses a specific use case, the results suggest that synthetic data could indeed be used as a surrogate when real data is scarce, incomplete or biased.
Virginia Fernandez, lead author and PhD student at the School of Biomedical Engineering & Imaging Sciences
The researchers designed brainSPADE with two variations: brainSPADE2D and brainSPADE3D, capable of producing 2D and 3D datasets respectively. Both versions offer users control over the type and severity of abnormalities present in the synthetic images. This allows researchers to tailor the training data to specific scenarios.
The potential benefits of brainSPADE are significant. By providing a means to train segmentation algorithms on a larger and more diverse dataset, brainSPADE has the potential to improve their accuracy in real-world applications. This could lead to more precise diagnoses, particularly for complex or rare medical conditions. The team is now looking at ways to further enhance the quality and realism of the synthetic images generated by brainSPADE, and also exploring the applicability of brainSPADE to other types of imaging besides MRI.