Development of Deep Learning Models for Unsupervised Labeling of Pathologic Features Apparent in Magnetic Resonance Imaging

MSc assignment

Successful development and application of supervised deep learning models in the past decade catapulted by the colossal quantities of manually labeled images readily available online -- primarily driven by giants like Google and Facebook. However, despite its promise, the success of this approach in medical applications remains elusive; our best models fail to identify as many as 20% of pathological imaging markers, and most are applicable only to a very limited set of clinically-relevant conditions. This precludes their clinical utility. The primary reason of this failure is the lack of sufficiently large, manually labeled images required by the antiquated supervised deep learning models. While part of this is lack can be attributed to the limited availability of clinical images, by far the largest reason is the limitations associated with labor intensive preprocessing an manual labeling.  

We have recently developed an unsupervised deep learning model to automatically identify and label an undetermined number of imaging features apparent in magnetic resonance (MR) images without the need for preprocessing and manual labeling. We have shown that this approach can successfully identify and segment white matter hyperintensities, a primary source of cognitive decline and dementia. Further development of this unsupervised model will remove the largest constraint facing existing machine- and deep learning based models -- labor intensive manual labeling. This, in turn, is likely to represent a major breakthrough in utility of these models in medical diagnosis by enabling development of large, labeled clinical data sets. Thus,

the goal of the project is to extend our unsupervised approach based on deep learning to label pathological features apparent in MR images, using clinical brain tumors and white matter hyperintensities as application cases.

Broadly, this project involves

  • further development and documenting mathematical and engineering underpinnings of the unsupervised MRI processing pipeline for 2D and 3D applications
  • benchmarking and testing multiple, popular deep learning models and evaluating their performance against the unsupervised model
  • application of the unsupervised model to clinical MRI images of white matter hyperintensities, brain tumors, and ischemic and hemorrhagic stroke


The project is suitable for master students from Biomedical Engineering and Electrical Engineering. Contact Elfi Hofmeijer ( and/or dr. Can Ozan Tan ( if you are interested in this assignment or would like to have more information.