Automatic Medical Image Segmentation using Machine Learning

State-of-the-art biophysical models are based on patient specific medical images, but there is a significant amount of manual labor at the pre-processing stage. The aim of the project is to design the machine learning approach for automatic segmentation of 3D medical images of the blood vessels in the brain.
Master

Cardiovascular diseases are burdening the healthcare systems and the costs are expected to rise in the years to come. Acute stroke alone is estimated to cost the European countries an overwhelming 40 billion annually. Although systemic risk factors have been associated with higher prevalence of cardiovascular diseases, the cause of a stroke, atherosclerotic plaques and defect balloon-shaped blood vessels in the brain (aneurysms), are focally distributed. This highlights the role of blood flow-induced wall shear stress (WSS) and its continuous role in vascular remodeling. Direct measurements of these stresses are difficult and medical image-based computational fluid dynamics (CFD) has been extensively used to study the 'patient-specific' local abnormal forces in search for a mechanistic biological link to disease initiation.
State-of-the-art biophysical models are based on patient specific medical images, but there is a significant amount of manual labor at the pre-processing stage, i.e., segmentation of the region of interest, to which the models will be
applied. In addition to that, the medical images contain quite a lot of noise leading to difficulties in distinguishing between arteries and surrounding tissue, not to mention image artefacts versus smaller vessels.

Goal

The aim of the current project is to design the machine learning approach for automatic segmentation of 3D medical images of the blood vessels in the brain. The impact of the developed tools will be significant, as it will enable large cohort studies of patient-specific cerebral blood flow.

Learning outcome

  • Good understanding of techniques for medical image processing and machine learning
  • In-depth understanding of the biomedical problem and tools for biophysical modelling
  • Understanding of dictionary learning and deep learning tools

Qualifications

  • Knowledge of image processing, machine learning
  • Tools for data representation learning like dictionary learning
  • Sparse coding

Supervisors

  • Valeriya Naumova
  • Per Magne Florvaag, MIND Simula
  • Kristian Valen-Sendstad, Comphy Simula