Nicolas Makaroff

This project enhances segmentation algorithms in convolutional neural networks (CNNs) using a new attention-gate mechanism based on Chan-Vese energy minimisation for improved precision. By imposing a constraint on segmentation via a partial differential equation (PDE), the approach offers more control over segmentation masks in a standard CNN architecture like U-Net. This results in competitive performance on binary segmentation while retaining crucial spatial information. The method's effectiveness is demonstrated using a database of brain images, showcasing its potential for accurate medical image segmentation.

This method uses geodesic distances and their geometric information for advanced image segmentation, particularly in data-oriented imaging applications. Combining image-based metrics with CNN-generated isotropic Riemannian metrics offers improved segmentation tailored to specific problems. In an example application, brain tumours are segmented as unit balls using geodesic distances calculated from CNN-output metric potentials, ensuring geometric and topological constraints. The results demonstrate that geodesic distance modules integrated into machine learning frameworks deliver state-of-the-art performance while maintaining geometric and topological properties.