Sep 2026· Medical Image Analysis· Vol 115, pp.
104316
· 0 citations· 48 references
Medicine
TL;DR
This study introduces FPGL, a framework that adapts INRs for medical segmentation by meta-learning a shared initialization by enabling both first-order (Reptile) and second-order (MAML) meta-learning routines.
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.
Overall, the studies reviewed here suggest that VAEs can serve as parts of larger systems, where they provide structured representations and help handle missing modalities.
Nabil Marzoug, M. Akhloufi· Journal of Imaging· 0 citations
Recent years have witnessed rapid progress in deep learning–based medical image segmentation. Existing approaches can be broadly categorized into discriminative models, which formulate segmentation as a pixel-wise classification problem, and generative models, which treat segmentation as a image conditioned mask genera...
Hao-Yu Cao, Yu-Xin Lin, Jia-Fa Lu et al.· IEEE Transactions on Image P...· 0 citations
While feed-forward segmentation networks perform well on large cohorts, they degrade in constrained clinical settings with few training samples and imaging artifacts. Human vision, by contrast, remains robust under these conditions, a property partly attributed to the recurrent and feedback connectivity of the visual c...
J. Marques, David Calhas, Arlindo L. Oliveira· 0 citations
A key challenge in medical image analysis is the scarcity of large annotated datasets for specific populations and diseases. As deep learning models rely heavily on labeled data, effective transfer learning strategies are needed to reduce the dependence on manual annotations. Self-supervised learning has emerged as a p...
Various task-specific medical segmentation models have demonstrated strong performance, their heavy reliance on extensive annotations and highly specialized architectures limits their versatility, scalability, and clinical applicability. Recently, adapting vision foundation models, particularly the segment anything mod...