Neural-Implicit Reconstruction Technique

• Neural-Implicit Reconstruction Technique (NIRT) is a method that represents unknown objects as continuous neural functions, ensuring physical consistency with measured data.
• It optimizes network parameters at reconstruction time in a scan-specific, unsupervised fashion, replacing pixel-wise processing with coordinate-based representations across modalities.
• NIRT integrates both implicit architectural regularization and explicit physical priors, though challenges remain in computational efficiency, scalability, and uncertainty quantification.

Neural-Implicit Reconstruction Technique (NIRT) denotes, in this literature, a class of inverse-problem methods in which the unknown image, volume, signal field, or surface is represented by a coordinate-based neural implicit function and fitted directly to measured data through an explicit forward model. Rather than reconstructing a pixel or voxel array directly, or learning a population-level measurement-to-image map, NIRT optimizes network parameters at reconstruction time so that the induced object is consistent with acquisition physics. Across CT, PET, MRI, photoacoustic tomography, sonar, and related dynamic settings, the defining pattern is a continuous coordinate-to-signal parameterization, scan-specific or instance-specific fitting, and measurement-domain data consistency, often combined with explicit regularization or additional implicit fields for auxiliary physical quantities such as coil sensitivities, motion, or quantitative parameter maps (Najaf et al., 18 Apr 2025, Moussaoui et al., 26 Mar 2025, Rao et al., 6 Jun 2025, Yao et al., 2024, Qadri et al., 2022).

1. Conceptual definition and scope

At its core, NIRT replaces the conventional reconstruction variable with a neural field. In sparse-view CT, this is written as an attenuation image represented by an implicit neural representation $f_\theta : \mathbb{R}^d \to \mathbb{R}$, optimized so that its projections match the sinogram under the CT forward operator; the cited CT work explicitly places this within the conceptual class of model-based iterative reconstruction (MBIR), with the key change being optimization over neural-network parameters rather than directly over pixels or voxels (Najaf et al., 18 Apr 2025). The same structural idea appears in PET, where the tracer activity image is modeled as a coordinate-based SIREN and fitted directly to measured sinograms with a Poisson likelihood, without external supervision (Moussaoui et al., 26 Mar 2025).

This distinguishes NIRT from two nearby categories that are often conflated with it. First, it is not neural post-processing: the network is not merely refining a reconstruction produced elsewhere, but is itself the reconstruction parameterization. Second, it is not necessarily a supervised inversion network trained across a population to map measurements to images. Several cited methods are explicitly per-scan, zero-shot, or unsupervised in the sense that they use only the current scan and the known physics model during optimization (Moussaoui et al., 26 Mar 2025, Feng et al., 2022, Yu et al., 28 Feb 2025).

A recurring implication is that NIRT functions as a physics-informed internal-learning framework. The forward model remains central, exactly as in classical inverse problems, but the admissible solution set is restricted to the nonlinear range of a coordinate-conditioned neural representation. This suggests that the main methodological shift is not abandonment of model-based reconstruction, but a reparameterization of the unknown object.

2. Canonical mathematical formulation

A generic NIRT problem is written as

$\min_{\theta} L(\mathcal{A}\{f_\theta\}, y),$

where $y$ denotes measured data, $f_\theta$ the implicit representation, and $\mathcal{A}$ the modality-specific forward operator (Najaf et al., 18 Apr 2025). In CT, one implementation evaluates the neural field on grid coordinates $\{z_i\}_{i=1}^n$, forms

$$x = \mathcal{E}\{f_\theta\} = (f_\theta(z_i))_{i=1}^n,$$

and then predicts measurements by

$\hat y = A\,\mathcal{E}\{f_\theta\},$

with $A$ the discrete projection matrix (Najaf et al., 18 Apr 2025). In MRI, the analogous role is played by the multi-coil encoding operator, typically involving sensitivity modulation, Fourier transform, and undersampling; in one scan-specific parallel MRI formulation, coil $j$ obeys

$\min_{\theta} L(\mathcal{A}\{f_\theta\}, y),$0

and the image $\min_{\theta} L(\mathcal{A}\{f_\theta\}, y),$1 is replaced by the grid-sampled output of a coordinate network $\min_{\theta} L(\mathcal{A}\{f_\theta\}, y),$2 (Feng et al., 2022).

The data fidelity term is modality-dependent. In PET, the forward model is

$\min_{\theta} L(\mathcal{A}\{f_\theta\}, y),$3

with a Poisson negative log-likelihood

$\min_{\theta} L(\mathcal{A}\{f_\theta\}, y),$4

which is then optimized over neural-network weights rather than voxel values (Moussaoui et al., 26 Mar 2025). In MRI, implementations vary between $\min_{\theta} L(\mathcal{A}\{f_\theta\}, y),$5-type and $\min_{\theta} L(\mathcal{A}\{f_\theta\}, y),$6-type k-space consistency terms. INR-CRISTAL, for example, writes a joint image/sensitivity objective using the multi-coil model

$\min_{\theta} L(\mathcal{A}\{f_\theta\}, y),$7

and in practice uses an $\min_{\theta} L(\mathcal{A}\{f_\theta\}, y),$8 data-consistency loss together with separate image and sensitivity regularizers (Rao et al., 6 Jun 2025).

A salient feature of these formulations is that the neural parameterization can be inserted at different levels of the inverse problem. Some methods parameterize the reconstructed image directly; others parameterize latent basis images, sensitivity fields, motion fields, or boundary fields. The common structure is still the same: a differentiable coordinate-to-quantity map is pushed through the forward physics, and reconstruction is performed by test-time optimization of the network parameters.

3. Representation design space

NIRT is not tied to a single neural architecture or even a single type of implicit quantity. The literature spans intensity fields, signed distance functions, semantic occupancy fields, k-space fields, sensitivity maps, motion fields, and quantitative parameter maps.

Coordinate inputs are usually low-dimensional spatial or spatiotemporal tuples such as $\min_{\theta} L(\mathcal{A}\{f_\theta\}, y),$9, $y$0, or $y$1. Hidden representations are then enriched by sinusoidal activations, Fourier features, or multiresolution hash encodings. SIREN-based constructions are prominent in PET and some MRI work because sinusoidal activations can represent both low- and high-frequency structure; one PET reconstruction uses a 5-layer SIREN with 256 features per layer and a SoftPlus output to enforce strictly positive activity values (Moussaoui et al., 26 Mar 2025). Hash-encoded coordinate MLPs are favored in several MRI and PACT settings because they provide high effective spatial frequency capacity with small decoders; the bilevel MRI method uses a trainable multiresolution hash encoder with a compact ReLU MLP (Yu et al., 28 Feb 2025), while LoREIN uses hash-encoded MLPs to generate spatial bases, coil maps, and quantitative parameter maps in 3D multi-parametric qMRI (Zhang et al., 10 Jun 2025).

The parameterized unknown may also be geometric rather than radiometric. NeuralCT models dynamic CT boundaries as a spatiotemporal signed distance function, with attenuation recovered from a soft occupancy transformation of the SDF rather than predicted directly as a scalar image field (Gupta et al., 2022). In forward-looking sonar reconstruction, the surface is likewise represented as the zero level set of an SDF, while a differentiable sonar renderer produces measurement predictions from that field (Qadri et al., 2022).

NIRT also extends naturally to auxiliary physical fields. INR-CRISTAL represents both the image and the coil sensitivity maps as coordinate-based neural functions and adds explicit regularization on the sensitivity fields (Rao et al., 6 Jun 2025). A flow-guided dynamic MRI method uses one INR for the complex-valued image sequence and a second INR for the optical flow field, coupling them through the optical flow equation $y$2 (Li et al., 21 Nov 2025).

4. Optimization, priors, and regularization

Although NIRT is often associated with “implicit regularization,” the surveyed methods show that implicit regularization is only one part of the design space. Many successful formulations combine neural parameterization with explicit penalties or constraints.

In some cases the regularization is mainly architectural. The PET SIREN reconstruction does not add an explicit hand-crafted penalty to the objective; regularization is described as arising from the neural parameterization itself, from SIREN’s representational bias, and from finite-iteration behavior (Moussaoui et al., 26 Mar 2025). Other methods are explicitly hybrid. The scan-specific parallel MRI method of 2022 uses a total-variation term together with the INR data-consistency loss,

$y$3

and additionally replaces predicted sampled k-space with acquired k-space at inference to enforce exact consistency at measured locations (Feng et al., 2022). Sparse-view PACT reconstruction uses $y$4 data fidelity plus TV regularization while optimizing the neural field representing the initial heat distribution (Yao et al., 2024). Dynamic sparse-view PACT extends this further with temporal TV and nuclear-norm penalties on the Casorati matrix of the reconstructed sequence (Xiao et al., 29 May 2025).

MRI work provides especially clear examples of explicit physical priors beyond the image itself. INR-CRISTAL adds a sensitivity-map regularizer

$y$5

and studies Fourier-$y$6, low-rank, and TV choices, finding TV on sensitivity maps most effective overall (Rao et al., 6 Jun 2025). Flow-guided dynamic MRI introduces an optical-flow regularizer, flow smoothness, and TV on both image and flow fields (Li et al., 21 Nov 2025). LoREIN couples a low-rank subspace prior with coordinate-based continuity and weighted nuclear norm minimization of quantitative maps (Zhang et al., 10 Jun 2025).

Optimization itself is a central issue because NIRT is typically solved at test time. The CT acceleration paper frames slow convergence as a major limitation and proposes two specific accelerations that preserve the core NIRT formulation: a preconditioned loss called filtered least squares and a nonlinear equality-constrained ADMM algorithm (Najaf et al., 18 Apr 2025). In PET, L-BFGS is reported as more stable and faster than first-order methods for the SIREN-based Poisson objective (Moussaoui et al., 26 Mar 2025). In scan-specific MRI, bilevel Bayesian optimization is used to select reconstruction hyperparameters such as learning rate and weight decay from held-out measured k-space, precisely because hyperparameter sensitivity is a practical barrier for instance-specific INRs (Yu et al., 28 Feb 2025).

These examples make two points clear. First, NIRT does not imply the absence of explicit priors. Second, the optimization algorithm is often as consequential as the representation.

5. Modalities and problem classes

The cited literature shows that NIRT is a modality-agnostic reconstruction pattern rather than a technique tied to a single scanner or signal model.

Domain	Implicit unknowns	Representative example
CT	Attenuation field $y$7 or spatiotemporal SDF	Sparse-view CT (Najaf et al., 18 Apr 2025), NeuralCT (Gupta et al., 2022)
PET	Activity field $y$8 with positivity constraint	End-to-end PET (Moussaoui et al., 26 Mar 2025)
Parallel MRI	Complex image, coil sensitivities, or both	INR-CRISTAL (Rao et al., 6 Jun 2025), scan-specific INR MRI (Feng et al., 2022)
Dynamic / multi-contrast MRI	Joint contrast stacks, respiratory-conditioned k-space, or image-plus-flow fields	Multi-contrast MRI (Niessen et al., 5 Sep 2025), ICoNIK (Spieker et al., 2023), flow-guided MRI (Li et al., 21 Nov 2025)
PACT	Initial pressure / heat distribution, static or dynamic spatiotemporal field	Sparse-view PACT (Yao et al., 2024), dynamic PACT (Xiao et al., 29 May 2025)
qMRI	Spatial bases, coil maps, quantitative parameter maps	LoREIN (Zhang et al., 10 Jun 2025)

Across these domains, NIRT can operate in image space or measurement space. Some methods represent the image intensity directly, while others represent raw k-space as a continuous function of acquisition coordinates. ICoNIK is a particularly clear example of the latter: it learns a neural implicit representation directly in k-space, conditioned on respiratory navigator value and k-space location, then synthesizes respiratory-resolved images by sampling the learned k-space function and applying inverse FFT (Spieker et al., 2023).

The same structural flexibility appears in dynamic and high-dimensional problems. One MRI method reconstructs all contrasts in a multi-contrast acquisition jointly by mapping 2D spatial coordinates to an $y$9-dimensional complex signal vector across contrasts, rather than reconstructing each contrast independently (Niessen et al., 5 Sep 2025). Another jointly reconstructs a spatiotemporal dynamic MRI sequence and its motion field using coupled INRs (Li et al., 21 Nov 2025). Dynamic PACT represents the entire image sequence as a single function of $f_\theta$0, allowing both sparse-view reconstruction and temporal interpolation by querying the trained field at unmeasured times (Xiao et al., 29 May 2025). This suggests that NIRT is especially well matched to settings where one wants a unified continuous representation across space, time, or contrast dimensions.

6. Limitations, misconceptions, and open directions

Several recurrent limitations appear across the literature. The most consistent is computational burden. Because many NIRT methods are scan-specific and solved by iterative test-time fitting, runtime is often measured in minutes rather than milliseconds. The PET SIREN reconstruction reaches best quality in only 10 iterations, but each iteration is costly (Moussaoui et al., 26 Mar 2025). The bilevel MRI framework is clinically plausible only because expensive hyperparameter optimization is moved offline and the resulting protocol-specific settings are reused for later scans (Yu et al., 28 Feb 2025). CT acceleration work explicitly targets the optimization bottleneck because thousands of iterations may otherwise be required (Najaf et al., 18 Apr 2025).

A second limitation is incomplete evidence for scalability and robustness. Multiple studies are confined to 2D or to realistic simulations rather than prospective clinical deployment. The PET example is 2D and simulation-based (Moussaoui et al., 26 Mar 2025). The sparse-view PACT study uses 2D ring-array formulations with homogeneous sound speed assumptions (Yao et al., 2024). Dynamic MRI work reports retrospective evaluations and notes that brightness-constancy assumptions may break under more complex dynamics (Li et al., 21 Nov 2025). These papers therefore support the feasibility of NIRT, but also leave open questions about 3D scaling, mismatch robustness, and deployment under real acquisition nonidealities.

A third limitation concerns uncertainty and identifiability. Several papers explicitly note that uncertainty quantification is not addressed, and that robustness to errors in attenuation, scatter, randoms, sensitivity maps, or motion modeling remains unresolved (Moussaoui et al., 26 Mar 2025, Rao et al., 6 Jun 2025). A plausible implication is that the compactness and continuity of neural parameterizations can regularize reconstructions effectively, but do not by themselves resolve ambiguity about which physically plausible explanation is correct.

The literature also corrects several common misconceptions. NIRT is not synonymous with supervised deep reconstruction; many representative methods are unsupervised, zero-shot, or scan-specific (Moussaoui et al., 26 Mar 2025, Yu et al., 28 Feb 2025). It is not limited to image intensity parameterization; sensitivity maps, motion fields, low-rank basis functions, and geometric boundary fields can all be represented implicitly (Rao et al., 6 Jun 2025, Li et al., 21 Nov 2025, Zhang et al., 10 Jun 2025, Gupta et al., 2022). Nor does continuous representation eliminate all discretization issues: even when the unknown is continuous, forward operators are often still numerically discretized, as in PACT where line or arc integrals are approximated by sampled points on the relevant wavefronts (Yao et al., 2024).

Current work points toward several open directions already identified within the cited papers: faster and better-conditioned optimization, stronger handling of forward-model mismatch, extension from 2D to realistic 3D clinical settings, explicit uncertainty modeling, and more systematic exploitation of auxiliary physical fields. A consistent theme is that NIRT is evolving from “implicit image fitting” toward broader physics-informed joint estimation, in which the neural field may represent not only the object of interest but also the latent structure that governs how that object is observed.