Deep neural networks for medical image segmentation

With the development of machine learning and especially deep learning, these have become efficient tools for analyzing medical images such as X-rays, computed tomography scans, or magnetic resonance imaging. Whether it is classification, segmentation, or registration of images, deep learning methods offer algorithmic solutions that significantly aid objective decision-making by healthcare professionals. Tasks such as segmentation are inherently imbalanced problems in medical imaging. In segmentation, the number of pixels/voxels in the background or organs is significantly higher than the number of pixels/voxels in some pathological structures. As part of this project, our goal is to develop new methods and approaches that not only take into account but also leverage the data imbalance and significantly improve the methods of segmenting anatomical structures.

The scientific objectives of the project:

1. Investigate new methods for processing medical images using the class-imbalanced nature of medical images. Several methods and approaches for processing medical images based on different principles and focused on various problematic areas of image processing, such as deep transfer learning, incomplete data, and imbalanced data, will be proposed within the project.
2. Propose image processing methods with a focus on medical images. Analyzing the training of deep networks under imbalanced data conditions. We propose approaches that can cope with domain shift.
3. Designing specific augmentation, loss functions, and their verification on multiple types of images such as X-rays, CT, and MRI, with a focus on the abdominal area. The design will consider the specificities of particular anatomical structures.
4. Propose new segmentation methods for anatomical structures in the abdominal area, such as organs, tumors, and cysts. This task represents a significant challenge due to the complexity of the abdominal area and the diversity of anatomical structures that need to be precisely identified. The new approaches will utilize the latest techniques in the field of computer vision and machine learning, including generative adversarial networks (GANs) and Capsule Networks.