How Far Can INRs Go? Cross-Domain Parameter-efficient INR-Based Semantic Segmentation for Brain MRI

What it claims

On cross-site 2D brain MRI segmentation, INR-based segmenters beat U-Net only in a specific regime: small parameter budgets with little or no augmentation. The comparison spans 5K to 1M parameters, with and without augmentation. As capacity grows and standard augmentation is applied, U-Net catches up and overtakes them. A layer-wise analysis of a meta-trained MetaSeg INR shows that segmentation-relevant structure sits in several hidden layers, not only the penultimate one MetaSeg reads. HierINRSeg fuses features from multiple layers. Averaged across all parameter budgets, it beats MetaSeg by 5.6 Dice points in-domain. The headline 8.2 out-of-domain figure is the mean of 7.2 and 9.2 on the two target sites.

How it works

The baseline, MetaSeg, uses a SIREN INR that maps a 2D coordinate to pixel intensity, plus an MLP segmentation head that reads the INR's penultimate-layer features. Training has two stages:

1. Meta-learning (MAML-style). For each training image, an inner loop adapts the INR on reconstruction loss and the head on segmentation loss. An outer loop updates the shared initialisation.
2. Head training. The meta-learned INR is fit to each training image. Penultimate features are extracted at every coordinate, and the head is trained further on ground-truth masks.

At inference, the INR is fine-tuned to reconstruct each test image before its features go to the frozen head. Every prediction therefore pays a per-image optimisation cost.

The layer analysis used PCA to 90% variance, then a shared UMAP over features taken before and after fitting. It found:

• First layer: low-level coordinate information, with no label grouping.
• Middle layers: stable under fitting, and they preserve a brain-shaped spatial-anatomical layout that transfers across images.
• Later layers: change substantially when fit to an image. They become label-discriminative and image-specific, and the anatomical layout disappears.

HierINRSeg aggregates features from multiple layers into the segmentation head instead of using only the last hidden layer. The fetched text ends before the exact fusion mechanism is described.

flowchart LR
  C[Coordinate x] --> INR[Meta-learned SIREN<br/>fine-tuned per test image]
  INR --> E[Early/middle layers<br/>stable, anatomical]
  INR --> L[Penultimate layer<br/>image-specific]
  L --> M[MetaSeg head]
  E --> F[Hierarchical fusion]
  L --> F
  F --> H[HierINRSeg head]

Evidence

• Setting: KKI is the source (in-domain) site. CMU is one of the out-of-domain target sites.
• Metric: foreground mean Dice.
• HierINRSeg vs MetaSeg: +5.6 in-domain, +7.2 and +9.2 on the two OOD sites, each averaged over all parameter budgets.
• Crossover: Figure 1 shows INR-based models ahead at small budgets and U-Net ahead at larger ones. The exact crossover point is not stated in the fetched text.
• Augmentation: the authors attribute the different response to augmentation to how augmentation changes features during INR fitting.

Limits

• The fetched text is truncated partway through Section 3. HierINRSeg's fusion design, the full result tables, dataset sizes, the second OOD site's name, and runtime figures are not visible.
• The method is 2D (coordinates in R²) and relies on spatially aligned images. The anatomical-structure finding is explicitly interpreted in light of that alignment.
• Inference requires test-time INR fine-tuning per image. Weigh this against the parameter savings if latency matters.
• The INR advantage is conditional. If you can afford roughly 1M parameters and standard augmentation, the paper's own results favour U-Net.
• The authors state that LLMs assisted with manuscript writing and code development.
• No code, weights or licence are given in the visible text. This is a v1 preprint dated 25 Sep 2026.