NOIR:
Neural Operator Mapping
for Implicit Representations

1 Department of Biomedical Engineering, University of Basel.   2 Harvard Medical School, Brigham and Women's Hospital.
Abstract
NOIR downstream tasks overview
Downstream tasks solved by NOIR. Segmentation, shape completion, synthesis, and image-to-image translation in 2D and 3D, with inherent robustness to resolution changes.

NOIR reframes core medical imaging tasks as operator learning between continuous function spaces, challenging the paradigm of discrete grid-based deep learning. Instead of operating on fixed pixel or voxel grids, NOIR embeds medical signals into shared Implicit Neural Representations (INRs) and learns a Neural Operator (NO) that maps between their latent modulations, enabling resolution-independent function-to-function transformations. We evaluate NOIR on segmentation, shape completion, image-to-image translation, and image synthesis across several public datasets (Shenzhen, OASIS-4, SkullBreak, fastMRI) and an in-house clinical dataset, in both 2D and 3D.

Neural Fields Neural Operators Medical Imaging Resolution Invariance
Method

Neural Operator Mapping for Implicit Representations

NOIR formulates each downstream task as an operator learning problem between continuous function spaces. Both input and output signals are embedded into shared INRs via meta-learning, producing compact latent codes. A Neural Operator then maps between these codes, decoupling the prediction from the input discretization.

NOIR architecture diagram
Architecture of NOIR. Discrete signals are embedded via INRs with shared parameters $\theta$ and signal-specific latents. A Neural Operator maps $z_\text{in} \to z_\text{out}$, enabling resolution-independent function-to-function mapping.
01

Challenges grid-based deep learning by questioning the fundamental choice of image representation.

02

Formulates downstream tasks as Neural Operator learning between INR latent spaces.

03

Applies to segmentation, shape completion, synthesis, and translation in 2D and 3D.

04

Empirically validates the $\varepsilon$-ReNO property: consistent predictions across various input discretizations.

Results

Segmentation & Shape Completion

Evaluated at 10%, 25%, and 100% of native resolution against U-Net, ViT, AttU-Net, and FNO. Grid-based methods degrade sharply at low resolution; NOIR maintains stable performance across all discretizations.

Qualitative segmentation and shape completion results
Qualitative results. Rows: Shenzhen binary segmentation, OASIS-4 multi-label brain MRI, SkullBreak 3D shape completion. Columns: input, ground truth, NOIR, and baselines.
Method Res 10% Res 25% Res 100%
DSCIoUDSCIoUDSCIoU
Shenzhen — Binary Segmentation
U-Net0.750.620.920.870.950.91
ViT0.910.840.930.870.940.90
AttU-Net0.010.010.740.640.950.91
FNO0.920.860.940.880.940.89
NOIR (Ours)0.930.880.940.880.940.88
OASIS-4 — Multi-label Brain MRI
U-Net0.400.290.790.660.950.90
ViT0.660.510.850.740.900.83
AttU-Net0.400.290.760.640.950.90
FNO0.650.510.840.740.910.83
NOIR (Ours)0.850.790.850.790.850.79
SkullBreak — 3D Shape Completion
3D U-Net0.710.560.900.830.960.93
3D ViT0.860.760.910.840.910.84
3D AttU-Net0.520.350.840.730.960.93
FNO0.740.590.870.780.900.82
NOIR (Ours)0.860.730.860.750.870.75

Image Synthesis & Translation

Qualitative synthesis and translation results
Qualitative results. Top: ultrasound synthesis from label maps. Bottom: MRI PD → T2 translation (fastMRI).
Method Res 25% Res 50% Res 100%
PSNRSSIMPSNRSSIMPSNRSSIM
Ultrasound Synthesis
U-Net17.510.8417.450.8417.540.89
ViT28.220.7828.310.7928.360.79
AttU-Net27.760.8728.290.8928.550.90
FNO27.370.7327.940.7528.150.76
DDPM26.350.6526.510.6526.910.64
NOIR (Ours)30.940.8630.940.8631.830.87
fastMRI — MRI PD → T2
U-Net21.680.6023.490.6824.550.72
ViT22.340.6222.540.6322.610.63
AttU-Net21.750.6023.260.6724.460.72
FNO22.610.5822.740.6022.820.60
DDPM22.660.6925.590.7625.450.74
NOIR (Ours)22.870.6122.870.6122.870.61
Theoretical Properties

NOIR is an $\varepsilon$-ReNO

A core desideratum for neural operators in scientific computing is consistency across discretizations: the same underlying operator should be recovered regardless of how the input function is sampled. Bartolucci et al. formalise this as the Representation Equivalent Neural Operator (ReNO) framework, and we show that NOIR empirically satisfies it.

A neural operator $\mathcal{T} : \mathcal{F} \to \mathcal{G}$ is a ReNO (Bartolucci et al.) if its discrete counterpart $\mathcal{T}^d$, acting on latent coefficients rather than grid values, is consistent across discretizations. Formally, the aliasing error between the continuous operator and the discrete pipeline must vanish:

Aliasing error
$$\varepsilon(\mathcal{T},\, \mathcal{T}^d) \;=\; \mathcal{T} \;-\; T_\Phi \circ \mathcal{T}^d \circ T_\Psi^\dagger \;\approx\; 0$$

$T_\Psi^\dagger$ maps a discrete signal to its latent code; $T_\Phi$ synthesises a continuous output from predicted latents. An $\varepsilon$-ReNO bounds this error by a small $\varepsilon > 0$, guaranteeing that two discretizations of the same input yield nearly identical predictions.

How NOIR instantiates the framework

In NOIR, $T_\Psi^\dagger$ is the INR inner-loop optimization (discrete signal $\to$ $z_\text{in}$) and $T_\Phi$ is the output INR evaluated at arbitrary resolution. Because $\mathcal{T}^d$ operates solely on $z_\text{in}$, it is decoupled from the input grid.

Empirical validation

We validate this on the Shenzhen dataset across five resolutions: $32^2$, $64^2$, $128^2$, $160^2$, $200^2$. For each test sample and each resolution pair $(r_i, r_j)$ we compute pairwise MSE between latent codes, and define $\varepsilon$ as the worst-case bound over the entire test set.

4.0 × 10⁻⁶
$\varepsilon_{z_\text{in}}$ — max pairwise MSE between input latents across all resolution pairs
2.2 × 10⁻⁵
$\varepsilon_{z_\text{out}}$ — max pairwise MSE between predicted output latents
0.93–0.94
Mean Dice across all resolutions $32^2$–$200^2$ — essentially flat
Dice distributions across resolutions
(a) Dice distributions from $32^2$ to $200^2$: mean stays at 0.93–0.94.
Segmentation at multiple resolutions
(b) Consistent segmentation predictions on the same test sample at different input resolutions.
Input latent MSE across resolutions
Input latent $z_\text{in}$. Pairwise MSE across resolutions — near zero, confirming resolution-independent encoding via $T_\Psi^\dagger$.
Output latent MSE across resolutions
Output latent $\hat{z}_\text{out}$. Pairwise MSE across resolutions — the NO produces nearly identical predictions regardless of the input discretization.
Supplementary — INR Reconstruction Quality

Signal Reconstruction

Output INR reconstructions vs. ground truth confirm the latent codes carry sufficient information for each task.

Shenzhen INR reconstructions
Shenzhen. Output INR reconstructions of lung segmentation masks. DSC: 0.98, IoU: 0.97.
OASIS-4 INR reconstructions
OASIS-4. Output INR reconstructions of multi-label brain segmentations. DSC: 0.91, IoU: 0.84.
fastMRI INR reconstructions
fastMRI. Output INR reconstructions of T2 MRI slices. PSNR: 29.09, SSIM: 0.71.
Ultrasound INR reconstructions
Ultrasound. Output INR reconstructions from latent codes. PSNR: 33.48, SSIM: 0.70.
Citation

BibTeX

@inproceedings{elhadramy2026noir,
  title     = {NOIR: Neural Operator Mapping for Implicit Representations},
  author    = {El Hadramy, Sidaty and Haouchine, Nazim and
               Wehrli, Michael and Cattin, Philippe C.},
  booktitle = {},
  year      = {2026}
}