Learning generalizable and semantic implicit neural representations for medical imaging
Abstract: Implicit neural representations (INRs) are new and emerging neural signal or data representations that typically represent a single signal or measurement by essentially overfitting the signal to an underlying multi-layer perceptron's parameters. However, INRs are sensitive to their underlying parameter initialization, which largely determines their convergence and the quality of the learned representation. In this talk, I will discuss how we can leverage common properties of natural images, such as underlying structures and low-frequency components, to learn efficient data-driven initialization schemes for implicit networks, allowing them to learn high-quality signal representations quickly. Building on the same premise, I will also present a joint, novel task-based initialization scheme for implicit networks, specifically for signal-translation tasks such as segmentation, known as MetaSeg, which received the 2025 MICCAI Best Paper Award. I will dive deeper into MetaSeg and how it performs on 2D and 3D Brain MRI segmentation.
- Tags
-