Leveraging expert knowledge for medical image segmentation

About the Project

Despite recent advances, deep learning methods for medical image segmentation can be hampered by limited availability of expertly labeled data. This project proposes a novel approach to enhance segmentation quality by incorporating constraints on invariance, topology, dimensions, position, count, shape, and other segmentation properties. The method will also handle weak annotations like scribbles and global properties such as unbiasedness.

Our key objective is to develop efficient algorithms that integrate deep learning with constraint optimization techniques. We will investigate penalty, barrier, and projection methods, the moving target method and feasibility-guaranteeing reparameterization. These techniques will be coupled with transformation-invariant network operators.

We are looking for one PhD student and one postdoc.

Requirements: strong programming, mathematical and research skills, prior experience with image processing and deep learning.

Computer Science (8)
Engineering (12)

Funding Notes

Funded by the Czech Science Foundation.

References

To learn about my research, see https://cmp.felk.cvut.cz/~kybic