IPDRecon: Physics-Driven Reconstruction

• IPDRecon is a family of inverse reconstruction methods defined by explicit forward models, iterative corrections, and structured priors to ensure physical consistency.
• It is applied across diverse modalities such as PET imaging, interior tomography, seismic interpolation, and electromagnetic scattering.
• The approach emphasizes practical enforcement of measurement physics and constraints to achieve more accurate and efficient reconstructions.

Searching arXiv for “IPDRecon” and closely related usages to ground the entry in current arXiv records. Searching arXiv for exact term and adjacent method names. IPDRecon denotes, in the supplied arXiv literature, a cluster of inverse-reconstruction ideas centered on combining measurement physics, explicit data consistency, and structured priors rather than a single universally standardized algorithm name. In one explicit usage, the term corresponds to a physically-based inverse rendering framework for PET image reconstruction built on Dr.Jit (Li et al., 27 Aug 2025). In adjacent usages, the same label is associated with differential phase-contrast interior tomography, projectional inverse-problem solvers, untrained deep-prior seismic interpolation, physics-driven neural inverse scattering, and time-imaging likelihood reconstruction in Cherenkov detectors (Cong et al., 2011, Dittmer et al., 2019, Liu et al., 2019, Du et al., 10 Dec 2025, Dzhygadlo et al., 2020). This suggests that IPDRecon is best understood as a family resemblance among reconstruction methods whose common elements are explicit forward models, iterative correction, and a controlled mechanism for restricting reconstructions to physically plausible or data-consistent solution sets.

1. Nomenclature and domain-specific usages

The supplied literature uses the label in several domain-specific ways. In PET, IPDRecon refers to a physically-based inverse rendering platform in which the voxel-wise tracer activity image is optimized by differentiating a hybrid Monte Carlo and analytical forward model with automatic differentiation (Li et al., 27 Aug 2025). In differential phase-contrast interior tomography, the term appears as a downstream descriptor for methods built on exact ROI reconstruction from truncated differential phase projection data under piecewise constant or piecewise polynomial priors (Cong et al., 2011). In seismic interpolation, the closely related designation is DSPRecon, an unsupervised U-net method that fits a single undersampled seismic record without external training data (Liu et al., 2019). In electromagnetic inverse scattering, the corresponding construction is an improved physics-driven neural network, abbreviated IPDNN, which the supplied text links to “IPDRecon / improved physics-driven neural network (IPDNN)” (Du et al., 10 Dec 2025).

Context	Designation in supplied literature	Defining mechanism
PET image reconstruction	IPDRecon / inverse rendering (IR) platform	Differentiable forward renderer with AD
Differential phase interior CT	Basis later methods such as IPDRecon build on	TV reconstruction from truncated differential projections
General inverse problems	Projectional Method (PM)	Alternating projections between prior and data-consistency sets
Seismic interpolation	DSPRecon	Single-record untrained U-net fitting
Electromagnetic inverse scattering	IPDNN	Physics-driven FCN with GLOW, dynamic subregion updates, transfer learning
PANDA Barrel DIRC PID	time-imaging reconstruction	Maximum-likelihood PDFs over photon pixel and time

A plausible implication is that the label is not tied to a single modality-specific implementation. Instead, it marks a reconstruction style in which forward operators, priors, and correction steps are kept explicit.

2. Mathematical lineage: interior tomography and projectional reconstruction

One major theoretical lineage begins with differential phase-contrast interior tomography. The imaging model uses the complex refractive index

$n(r)=1-\delta(r)+i\beta(r),$

with measured data given by derivatives of the phase projection,

$\frac{\partial}{\partial s} R[\delta](s,\theta).$

The interior problem is to recover $\delta(r)$ on an ROI $\Omega_a$ from truncated differential phase projection data known only for $s\in(-a,a)$. The paper proves a uniqueness result: if the refractive index image is piecewise polynomial in $\Omega_{2A}$ and another image has the same truncated differential phase projection data on the ROI, then the two images are equal. For the piecewise-constant case, the correct reconstruction is the minimum-TV solution among all images matching the truncated differential phase data, and the practical solver is a split Bregman algorithm with a least-squares subproblem, $\ell_1$ shrinkage steps, and Bregman multiplier updates (Cong et al., 2011).

A second lineage is the projectional formalization of inverse problems. "A Projectional Ansatz to Reconstruction" casts reconstruction as the search for a point in the intersection of a prior set $U$ and a data-consistency set

$$V(A,y^\delta,\delta):=\{x\in X:\|Ax-y^\delta\|=\delta\},$$

or, in practice, its convex relaxation

$$\overline V(A,y^\delta,\delta):=\{x\in X:\|Ax-y^\delta\|\le \delta\}.$$

The Projectional Method alternates between $\frac{\partial}{\partial s} R[\delta](s,\theta).$0 and $\frac{\partial}{\partial s} R[\delta](s,\theta).$1, invoking von Neumann alternating projections. The projection onto the valid-solution set is defined by

$\frac{\partial}{\partial s} R[\delta](s,\theta).$2

with solution

$\frac{\partial}{\partial s} R[\delta](s,\theta).$3

where $\frac{\partial}{\partial s} R[\delta](s,\theta).$4 is chosen so that $\frac{\partial}{\partial s} R[\delta](s,\theta).$5. The same paper derives the Jacobian of this projection layer via the implicit function theorem, making $\frac{\partial}{\partial s} R[\delta](s,\theta).$6 a differentiable data-consistency-enforcing layer for end-to-end learning. It also interprets RED as a relaxation of the projectional method and proposes the discrepancy-aware DIP loss

$\frac{\partial}{\partial s} R[\delta](s,\theta).$7

which is intended to stop the late-stage degradation of vanilla Deep Image Prior reconstructions by targeting the residual level $\frac{\partial}{\partial s} R[\delta](s,\theta).$8 rather than zero residual (Dittmer et al., 2019).

Taken together, these two works establish a recurrent IPDRecon-style principle: reconstruction is not merely denoising or fitting, but constrained recovery on a measurement-defined manifold, with regularity enforced either by TV-type sparsity or by an explicit prior projection.

3. Implicit-prior and single-instance reconstruction: DSPRecon

In seismic interpolation, the supplied literature identifies a closely related single-instance paradigm under the name DSPRecon. The reconstruction problem is posed on a complete seismic dataset $\frac{\partial}{\partial s} R[\delta](s,\theta).$9 and an observed undersampled dataset $\delta(r)$0 using a masking operator

$\delta(r)$1

The network output is defined by

$\delta(r)$2

subject to the observed-trace constraint

$\delta(r)$3

and the optimization objective

$\delta(r)$4

The method is explicitly unsupervised: there is no training or validation dataset, no pre-trained model, and only the single corrupted seismic record is used. The input is fixed random noise $\delta(r)$5, uniformly sampled between 0 and 0.1 and having the same size as the seismic data. At each iteration, Gaussian perturbation with $\delta(r)$6 is added as noise-based regularization. Optimization uses Adam with learning rate 0.001 for 3000 iterations (Liu et al., 2019).

The U-net architecture is specified as an encoder-decoder with skip connections, convolution layers with $\delta(r)$7 filters, batch normalization and LeakyReLU after each convolution, encoder filter counts 16, 32, 64, 128, 128, downsampling via $\delta(r)$8 stride-2 convolution, bilinear interpolation upsampling in the decoder, decoder filter counts 128, 128, 64, 32, 16, and skip connections that copy and add encoder features to decoder features. The method is stated to handle both irregularly missing traces and regularly missing traces. The supplied implementation details include an NVIDIA GTX 1080 Ti GPU, runtime about 7 minutes, normalization of seismic data to $\delta(r)$9, and the empirical observation that SNR stabilizes around 3000 iterations (Liu et al., 2019).

Experiment	DSPRecon	Comparator
Irregular sampling, synthetic pre-stack data	32.68 dB	SSA: 19.11 dB
Irregular sampling, post-stack data	33.09 dB	SSA: 24.27 dB
Regular sampling, field data	35.91 dB	de-aliased Cadzow: 15.32 dB

The comparison is framed against classical rank-reduction methods. SSA is described as classical for irregular trace reconstruction, and de-aliased Cadzow for regular missing traces. DSPRecon replaces the rank prior with a deep seismic prior encoded by the CNN and avoids manual rank tuning. This places it close to the projectional and implicit-prior strand of IPDRecon: the model learns from the single corrupted instance, while the masking operator preserves exact fidelity on measured entries.

4. PET inverse rendering as an explicit IPDRecon formulation

The most explicit use of IPDRecon in the supplied material is the physically-based inverse rendering framework for PET image reconstruction. The forward model is written as

$\Omega_a$0

where $\Omega_a$1 is the voxelized tracer activity image, $\Omega_a$2 is the coincidence sinogram, $\Omega_a$3 is the system matrix, and $\Omega_a$4 denotes random and scattered coincidences. The measured data are modeled as independent Poisson random variables, and the paper also gives the standard MLEM multiplicative update

$\Omega_a$5

The framework is described as the first PET reconstruction platform built on Dr.Jit and treats the emission image as unknown scene parameters, the measured sinogram as the observation, and the PET system as a differentiable forward renderer. Reconstruction is posed as

$\Omega_a$6

with iterative update

$\Omega_a$7

where gradients are obtained by automatic differentiation rather than hand-derived PET update equations (Li et al., 27 Aug 2025).

The forward renderer is hybrid. It combines Sobol Monte Carlo sampling within detector crystals,

$\Omega_a$8

with Gaussian perturbation for PSF, analytical line integrals along sampled sub-LORs,

$\Omega_a$9

optional TOF modeling by a Gaussian kernel along the LOR, and normalization and attenuation corrections through

$s\in(-a,a)$0

The contribution function is

$s\in(-a,a)$1

The paper emphasizes that automatic differentiation makes it feasible to include complex physics such as Compton scattering, positron range blurring, and annihilation photon acollinearity without manually deriving new update equations (Li et al., 27 Aug 2025).

The evaluation covers an ultra-micro Derenzo phantom, a 3D Hoffman brain phantom, and human brain PET. In the Derenzo study, IR resolved all hot spots of 1.0 mm and larger, consistently had higher SNR than CASToR across all spot sizes, and had 12.33% higher PVR at 1.0 mm, although CASToR had slightly better PVR for diameters $s\in(-a,a)$2 mm. In the Hoffman phantom, Table I reports COV of 8.013% for CASToR and 7.842% for IR, and SOR of $s\in(-a,a)$3 for CASToR and $s\in(-a,a)$4 for IR. In the human study, Table II reports COV of 13.23% for Biograph mCT and 17.06% for IR, SUVR of $s\in(-a,a)$5 for Biograph mCT and $s\in(-a,a)$6 for IR, and GWR of $s\in(-a,a)$7 for Biograph mCT and $s\in(-a,a)$8 for IR, including a reported 9.97% higher SUVR in the right hippocampus and 3.52% higher GWR. Runtime was also reported to be faster than CASToR: 1440.3 s versus 6373.0 s for the Hoffman phantom, about 4.5× faster, and 243.7 s versus 2374.0 s for the hotspot phantom, about 9.75× faster (Li et al., 27 Aug 2025).

Within the supplied corpus, this PET formulation is the clearest instance in which IPDRecon names a complete reconstruction platform rather than a general style.

5. Physics-driven neural inverse scattering

A second explicit modern usage links IPDRecon to the improved physics-driven neural network for electromagnetic inverse scattering. The problem is two-dimensional inverse scattering for TM, $s\in(-a,a)$9-polarized illumination, where the goal is to reconstruct the spatial relative permittivity distribution $\Omega_{2A}$0 from measured scattered fields on an observation surface. The physics is written through the state and data equations

$\Omega_{2A}$1

$\Omega_{2A}$2

with

$\Omega_{2A}$3

The forward solve is carried out by the method of moments. The reconstruction network is a fully connected network with two channels corresponding to the real and imaginary parts of the permittivity, and the total loss is

$\Omega_{2A}$4

where

$\Omega_{2A}$5

$\Omega_{2A}$6

and

$\Omega_{2A}$7

Optimization uses Adam with initial learning rate $\Omega_{2A}$8 (Du et al., 10 Dec 2025).

The principal architectural contribution is the Gaussian-localized oscillation-suppressing window activation,

$\Omega_{2A}$9

with derivative

$\ell_1$0

The intended behavior is linear growth for small inputs and exponential decay for large inputs, with the stated effects of suppressing abnormal outliers and oscillations, stabilizing training, and permitting a lightweight FCN. The paper reports training memory reduction from 264.49 MB for PDNN to 16.03 MB for the FCN/GLOW version (Du et al., 10 Dec 2025).

The second major component is dynamic scatter subregion identification. An initial estimate is obtained using a data-driven U-Net solver, thresholding forms a binary map through

$\ell_1$1

where $\ell_1$2 and $\ell_1$3 are computed from the smallest 30% of relative-permittivity values. Thresholding is repeated every 100 iterations, updates begin only after iteration $\ell_1$4, and the first update uses $\ell_1$5. The paper also incorporates transfer learning by pretraining on a sound object and fine-tuning on a defective object; in one example, the pretrained model reveals the defect after only 10 iterations. Validation includes simulated and Fresnel Institute experimental data, relative prediction error as the main quantitative metric, noise levels of 10%, 30%, and 50%, and reconstruction that remains effective even with only 4 transmitters in the pretrained experimental setting, although low-contrast scatterers degrade somewhat (Du et al., 10 Dec 2025).

This usage preserves the defining IPDRecon pattern: the network is not a black-box regressor but an inverse-solver parameterization constrained by a forward model, physically motivated penalties, and iterative domain refinement.

6. Adjacent formulations and recurrent design patterns

Other supplied works reinforce the same reconstruction logic in different observables. In the PANDA Barrel DIRC, time-imaging reconstruction performs particle identification from photon hit position and arrival time through the likelihood

$\ell_1$6

The signal PDF $\ell_1$7 may be constructed from data, detailed simulation, or analytically. In the analytical form,

$\ell_1$8

and a Geant4-generated lookup table is used in the full PANDA geometry to store possible photon directions at the radiator exit and propagation time in the expansion volume. The detector is designed for $\ell_1$9 separation of at least 3 standard deviations up to 3.5 GeV/$U$0 over polar angles $U$1 to $U$2, with single-photon position resolution of about 2 mm and time precision of about 100 ps. Reported $U$3 separation values are $U$4 standard deviations for analytical PDFs, $U$5 for simulated PDFs, and $U$6 for geometrical reconstruction (Dzhygadlo et al., 2020).

In nonlinear transient impact reconstruction, the improved ReConNN replaces the earlier CNN plus CWGAN structure with AVAE-based feature extraction, latent-space interpolation, and CGAN-based visual enhancement. The latent feature dimension is set to 1 for the transient impact case, the images are resized to $U$7, training is run for 150 epochs, and each interval is interpolated by 9 new values. The paper reports that the preferred case uses 563 samples, compared with 22,000 samples in Ref. [1] and 6,055 samples in Ref. [2]. For AVAE versus VAE, the reported numbers are 21.08 dB, 89.21%, and $U$8 for PSNR, SSIM, and MSE, compared with 19.93 dB, 84.64%, and $U$9 for VAE. The supplied discussion interprets this as improved accuracy, efficiency, and cost for reconstructing time-dependent ordered images in impact problems (Li et al., 2019).

Taken together, these works suggest several recurrent IPDRecon characteristics. First, the measurement process is retained explicitly, whether as differential phase projections, masked seismic traces, PET coincidence transport, MoM-based scattering physics, or photon arrival-time PDFs. Second, priors are encoded structurally rather than only statistically: TV, projection onto a plausibility set, network architecture bias, bounded-permittivity penalties, or latent-manifold interpolation. Third, optimization is iterative and correction-driven, with alternating projections, Adam, split Bregman, MLEM-style updates, or direct likelihood maximization. Fourth, several formulations explicitly address the failure mode of unconstrained fitting: DIP overfitting in projectional reconstruction, weak-scatterer omission in inverse scattering, rank-tuning difficulty in seismic interpolation, and noise-resolution trade-offs in PET (Dittmer et al., 2019, Liu et al., 2019, Li et al., 27 Aug 2025, Du et al., 10 Dec 2025).

In that sense, IPDRecon is best viewed not as a single fixed algorithm, but as a reconstruction idiom: physically specified forward modeling combined with iterative enforcement of admissibility, consistency, and prior structure.