Researchers have developed a method for expediting the creation and customization of templates used in medical image analysis to aid in disease diagnosis.
The new technique can be used to crunch datasets of patients’ medical images and capture structural relationships that may indicate disease progression. In many cases, analysis necessitates using a common image template known as an “atlas,” an average representation of a given patient population. Atlases, for example, can identify clinically significant changes in brain structures over time.
Creating a medical imaging analysis template is time-consuming and labor-intensive, often taking days or weeks, especially when 3D brain scans are used. Researchers frequently download publicly available atlases created by research groups to save time. However, these do not fully capture the diversity of individual datasets or specific subpopulations, such as those from new diseases or children. Finally, the atlas cannot be smoothly mapped onto outlier images, resulting in subpar results.
The researchers describe an automated machine learning model that creates “conditional” atlases based on patient characteristics such as age, gender, and disease. The medical image analysis model can create atlases from patient subpopulations that may be completely missing by leveraging shared information across an entire dataset. More atlases are needed in the world. Many medical image analysis models rely on atlases. This method can create a lot more of them as well as conditional ones.