FPGL: A meta-learned implicit neural representation framework for medical image segmentation.
Authors
Affiliations (5)
Affiliations (5)
- Department of Electrical and Computer Engineering, Rice University, Houston, TX, USA. Electronic address: [email protected].
- Department of Electrical and Computer Engineering, Rice University, Houston, TX, USA.
- The University of Texas MD Anderson Cancer Center, Houston, TX, USA. Electronic address: [email protected].
- Department of Electrical and Computer Engineering, Rice University, Houston, TX, USA. Electronic address: [email protected].
- Department of Electrical and Computer Engineering, Rice University, Houston, TX, USA. Electronic address: [email protected].
Abstract
Modern state-of-the-art deep learning architectures for medical image segmentation rely strictly on feed-forward passes over dense pixel/voxel grids, which scale poorly to large signals. Implicit neural representations (INRs) offer a lightweight, continuous alternative to raw grids, but are traditionally signal-specific and lack the semantic coherence needed for dataset-level predictive tasks. In this study, we introduce FPGL, a framework that adapts INRs for medical segmentation by meta-learning a shared initialization. By jointly fitting scans and segmentation maps across a training dataset, FPGL can segment novel subjects simply by fine-tuning on their raw test-time observations. We expand on our preliminary MICCAI study of this work in two key ways. First, we enable both first-order (Reptile) and second-order (MAML) meta-learning routines. Second, we introduce a generalized formulation that allows for segmentation from indirect measurements of a scan, such as tomographic projections and demonstrate several proofs-of-concept experiments across various forward and inverse tasks. We evaluate FPGL on a well-aligned brain 2D/3D MRI dataset and a more heterogeneous abdominal 2D CT dataset. On the MRI data, FPGL matches U-Net baselines. However, performance drops significantly on the CT dataset, revealing a sensitivity to spatial misalignments unexplored in our previous MRI-only study. Interestingly, performing CT segmentation directly from the scans and indirectly from sparse synthetic tomographic projections yields remarkably similar performance, suggesting that FPGL's current optimization design is not yet fully exploiting all details in a scan. We conduct additional studies and reveal key insights on the task- and optimizer-specific behavior of FPGL. By outlining both its promise and its current limitations, this study establishes a strong foundation for advancing the exciting new field of INR-based medical image segmentation. Our codebase is available at https://github.com/kushalvyas/FPGL.