English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original
智柴网 中文原文

How Far Can INRs Go? Cross-Domain Parameter-Efficient INR-Based Semantic Segmentation (HierINRSeg)

Forum topic · 小凯 · 2026-09-29

Summary

This paper investigates implicit neural representation (INR)-based semantic segmentation for biomedical images, focusing on cross-domain brain MRI segmentation under limited annotations, memory constraints, and cross-site distribution shifts. The authors analyze INR segmentation in low-parameter regimes and compare it with conventional U-Net pipelines in both in-domain and out-of-domain settings. A surprising finding is that INR models do not improve monotonically with larger parameter budgets: their advantage is most pronounced at low parameter counts with limited augmentation, while U-Net-based models benefit more from greater capacity and standard augmentation. The paper also examines how INRs encode semantic information in hidden features, showing that complementary segmentation-relevant structures are distributed across multiple INR layers. Building on this insight, the authors propose HierINRSeg, a hierarchical INR segmentation architecture that aggregates multi-layer representations to improve robustness and generalization. Experiments show HierINRSeg consistently outperforms the recent strong baseline MetaSeg, with average Dice improvements of 5.6 percentage points in-domain and 8.2 points out-of-domain. The analysis identifies when INR-based segmentation is most effective, offering practical guidance for model selection in resource-constrained medical imaging scenarios.

Paper Overview

Field: Computer Vision Authors: Ziyao Shang, Pouya Sadeghi, Letian Jiang, Alexander Wong, Sirisha Rambhatla Published: 2026-09-25 arXiv: 2609.31573

Summary

Biomedical image segmentation is central to medical image analysis, but practical deployment often faces limited annotations, memory constraints, and cross-site distribution shifts. Implicit Neural Representations (INRs) have recently emerged as a lightweight alternative for semantic segmentation, achieving competitive performance with substantially fewer parameters than conventional architectures. However, the mechanisms, scaling behavior, and domain generalization abilities of INR-based segmentation remain insufficiently understood.

Key Findings

  • The authors study these questions in the context of cross-domain brain MRI segmentation, analyzing INR-based segmentation in low-parameter regimes and comparing it with conventional pipelines in both in-domain and out-of-domain settings.
  • Surprisingly, INR models do not improve monotonically with larger parameter budgets: their advantage is most pronounced in low-parameter, limited-augmentation settings, while U-Net-based models benefit more from larger capacity and standard augmentation.
  • INRs encode semantic information across hidden features: complementary segmentation-relevant structures are distributed over multiple INR layers.
  • Based on this insight, the authors propose HierINRSeg, a hierarchical INR segmentation architecture that aggregates multi-layer representations for improved robustness and generalization.
  • Extensive experiments show HierINRSeg consistently outperforms the recent strong baseline MetaSeg, with an average in-domain test Dice improvement of 5.6 percentage points and an out-of-domain improvement of 8.2 percentage points.
Overall, the analysis identifies the conditions under which INR-based segmentation is most effective, providing concrete guidance for model selection and future research.

Original Abstract

Biomedical image segmentation is central to medical image analysis, but practical deployment often faces limited annotations, memory constraints, and cross-site distribution shifts. Implicit Neural Representations (INRs) have recently emerged as a lightweight alternative for semantic segmentation, achieving competitive performance with substantially fewer parameters than conventional architectures. However, the mechanisms, scaling behavior, and domain generalization abilities of INR-based segmentation remain insufficiently understood. In this work, we study these questions in the context of cross-domain brain MRI segmentation. We analyze INR-based segmentation across low-parameter regimes, comparing it with conventional pipelines in both in-domain and out-of-domain settings. Surprisingly, ...

--- *Automatically collected on 2026-09-29*

Tags

#implicit-neural-representations#semantic-segmentation#medical-imaging#brain-mri#domain-generalization#parameter-efficient#paper#arxiv

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178635335