Disputation: Deep Regression and Segmentation for Medical Inference: from Large-Scale Magnetic Resonance Imaging

  • Date:
  • Doctoral student: Doctoral student Taro Langner
  • Contact person: Joel Kullberg (huvudhandledare)
  • Disputation

Taro Langner defends his thesis Deep Regression and Segmentation for Medical Inference: from Large-Scale Magnetic Resonance Imaging.


Please attend the public defence through Zoom: https://uu-se.zoom.us/j/64556018436


Abstract [en]

Large-scale studies, such as UK Biobank, acquire medical imaging data for thousands of participants. With magnetic resonance imaging (MRI), comprehensive representations of human anatomy can be provided for non-invasive assessments of health-related conditions, body composition, organ volumes, and more. The sheer quantity of resulting image data itself poses a challenge, however, as manual processing and evaluation at the given scale is typically no longer feasible. 

For automated image analysis, machine learning techniques involving deep learning with convolutional neural networks have established state-of-the-art results in recent years. These systems can perform a multitude of tasks on medical image data, such as predicting measurements, classifying certain conditions, and enhancing image quality. The overall aim of this thesis was to explore the potential of deep learning for automated analysis of large-scale MRI derived from several studies.

Fully-convolutional networks for semantic segmentation were adapted and evaluated for the automated delineation and quantification of adipose tissue depots and abdominal organs.

As an alternative approach, convolutional neural networks were trained for image-based, deep regression to predict numerical values corresponding to measurements and abstract properties such as age or health states directly. The numerical values resulting from this regression approach are not easily explainable, as no intermediate segmentation masks are generated. For an interpretation of the decision criteria learned by the networks, aggregated saliency analysis was proposed as a visualization technique for relevant anatomical structures in thousands of co-aligned subjects. Additionally, methods for uncertainty quantification were adapted to provide individual confidence intervals along with each prediction.

By exploring different configurations and developing fast and effective strategies with these two methodologies, several software tools were implemented that can robustly predict measurements for thousands of UK Biobank subjects within hours, with no requirement for human guidance or intervention.

Link to the doctoral thesis in DiVA.

Events for Medicine and Pharmacy