Superpixel uncertainty modeling

Superpixel segmentation of medical images could be useful for the efficient detection of clinically relevant attributes such as melanomas or tumors. So far, there is no work on predicting the uncertainty of the segmentation masks that the models generate. The uncertainty is, however, important to understand how certain the model is about a specific region in the image. High model uncertainty could mean that a second opinion from a medical doctor is necessary to avoid incorrect and potentially harmful conclusions. This project presents a new method for uncertainty prediction of superpixel segmentation masks.

Goal

The goal of this project is to work on a new method for uncertainty prediction of superpixels for image segmentation. The experiments will be conducted on various available medical image datasets.

Learning outcome

Interdisciplinary research, machine learning and AI, performing research with advanced deep learning algorithms for real world applications.

Qualifications

Hard working, motivated, interested in learning

Supervisors

• Pål Halvorsen
• Michael Riegler
• Thu Nguyen, Simula, thu@simula.no
• Andrea Storås, Simula, andrea@simula.no