Machine learning creates image ‘atlas' to improve disease diagnoses

Massachusetts Institute of Technology researchers are using the power of machine learning to create customized templates radiologists lean on when diagnosing diseases.

Often times clinicians will analyze datasets of patients' medical scans to identify a relationship indicating the progression of diseases. This typically requires a template or “atlas” representative of a patient population which, when compared to the scan in question, highlights important differences.

The problem, according to Adrian Dalca, a faculty member in radiology at Harvard Medical School and Massachusetts General Hospital, is that building such templates can take weeks, especially when using three-dimensional scans; even then the final product can leave out crucial datapoints.

"The world needs more atlases," Dalca, a former postdoc in MIT’s Computer Science and Artificial Intelligence Laboratory, said in a university news story. "Atlases are central to many medical image analyses. This method can build a lot more of them and build conditional ones as well."

Dalca and colleagues at MIT combined two neural networks: one that learns an atlas after each image in a dataset is compared to the patient’s original scan, and another that aligns that atlas to images in a larger set. The approach eventually spits out a conditional atlas based on a patient’s specific attributes such as age, sex and disease.

Essentially, these conditional templates have learned how a person’s age, for example, correlates to structural variations in scans across all images in a dataset. It can do so even if there is little to no data on a specific patient population.

The researchers hope clinicians can use their model to build institution-specific atlases based on their own, potentially small datasets.

Dalca and colleagues are already working with a team at Massachusetts General to create conditional atlases from pediatric brain scans, which are not readily available. Their ultimate goal, however, is to establish a system that could generate an atlas for any subpopulation.

“Researchers could log into a webpage, input an age, sex, diseases and other parameters, and get an on-demand conditional atlas,” Dalca explained. “That would be wonderful, because everyone can refer to this one function as a single universal atlas reference.”

The full paper detailing these findings is set to be presented at the Conference on Neural Information Processing Systems beginning Dec. 8.