REVIEW 4 major objections 4 minor 25 references
Accelerated Optimization of Implicit Neural Representations for CT Reconstruction
T0 review · 4 major / 4 minor · reviewed 2026-08-16 · deepseek-v4-flash
Pith's one-line read Two optimizer changes—a filtered least-squares loss and ADMM—speed INR-based CT reconstruction to TV-level quality.
desk verdict A solid, incremental optimization recipe for INR-based CT: the conditioning analysis is clean, but the ADMM speed-up claim hinges on an iteration count that omits the INR Adam steps, so wall-clock comparison is needed before believing the headline acceleration. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing objects are the filtered least-squares (FLS) loss and the nonlinear constrained ADMM split. FLS is the quadratic loss $L_{\mathrm{FLS}}(z,y)=(z-y)^\top F(z-y)$ with $F$ a positive-semidefinite ramp-filter convolution, equal to $\|F^{1/2}(z-y)\|^2$; it acts as a preconditioner on the inner matrix $P^\top F P$ and improves the conditioning of $Q^\top P^\top F P Q$ relative to $Q^\top P^\top P Q$. The ADMM machinery rewrites the fitting problem as $\min \tfrac{1}{2}\|P x-y\|^2$ subject to $x=\mathcal{E}\{f_\theta\}$, then alternates a CGLS $x$-update, an INR $\theta$-update trained in pixel space without forward/back projections, and a multiplier update. Inexact solves with 50 CGLS and 50 Adam steps keep each outer iteration cheap. A linearized analysis treats the INR feature matrix $Q$ as fixed and reduces the convergence-rate question to Gram matrix condition numbers, which is what the empirical condition-number ratios in Table 1 measure.
What would settle it
Run the same sparse-view breast CT reconstructions to a fixed target image MSE on identical hardware while measuring wall-clock time and counting every Adam and CGLS inner iteration; if ADMM is not faster to target MSE than the filtered least-squares loss, the ADMM acceleration claim fails.
Extended reading notes
Core claim
The central claim is that the bottleneck in INR-based CT reconstruction is the ill-conditioning of the projection matrix, not the INR architecture. Even when the INR feature matrix $Q$ is well-conditioned, the Gram matrix $Q^\top P^\top P Q$ inherits ill-conditioning from $P^\top P$. The paper proposes two cures: replacing the least-squares loss with filtered least squares, $L_{\mathrm{FLS}}(z,y) = (z-y)^\top F (z-y)$, where $F$ is a positive-semidefinite ramp-filter matrix from filtered back-projection, and solving the constrained problem $\min_\theta \tfrac{1}{2}\|P x-y\|^2$ subject to $x = \mathcal{E}\{f_\theta\}$ by ADMM. At initialization the filtered loss reduces the Gram condition number by one to three orders of magnitude across ReLU, SIREN, and hash-encoding architectures. On a simulated 128-view sparse-view breast CT phantom, both FLS and ADMM produce reconstructions visually similar to a total-variation-regularized baseline with smaller or comparable MSE within 1000 counted iterations, whereas plain least-squares training is much slower and worse; ADMM achieves the lowest final MSE in all three architectures.
Load-bearing premise
The speed comparison counts one iteration as a single matrix-vector product with $P$ and $P^\top$; for LS and FLS that is one gradient step, but for ADMM it is one CGLS step, and the 50 Adam updates on the INR inside each ADMM outer iteration are not counted, so if wall-clock time is the real metric the ADMM speedup may be smaller than reported.
Editorial extensions
If this is right
- FLS and ADMM bring INR-based sparse-view CT into the same iteration budget as classical iterative reconstruction: enough to match or beat a TV-regularized baseline in 1000 counted iterations.
- The gain transfers across three structurally different INR architectures (ReLU with Fourier features, SIREN, and hash encoding), so the acceleration is not tied to one network design.
- ADMM gives the lowest final image MSE in every architecture tested, while FLS gives the fastest early error reduction for two of the three.
- Because the FLS improvement is a preconditioning of $P^\top P$, it should stack with other conditioning fixes for the INR feature matrix $Q$.
- The ADMM formulation moves INR training into pixel space, so each inner INR step avoids costly forward and back projections, which is the main source of its per-iteration efficiency.
Reading between the lines
- A natural next step is to benchmark wall-clock time with all inner ADMM Adam updates counted; the paper's counted-iteration convention may make the ADMM gain look larger than a fair timing would.
- Combining FLS with ADMM—for example, using a filtered loss for the pixel-space INR subproblem—is a direct test that could compound the two accelerations.
- The same conditioning argument should apply to other linear inverse problems where an FBP-like preconditioner exists, such as sparse-view MRI or deblurring, and the paper's approach could be ported there.
- Since the constraint $x=\mathcal{E}\{f_\theta\}$ is nonlinear, ADMM convergence is not guaranteed; monitoring residuals is a stopgap, and local convergence analysis would settle when the method is safe to use.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes two modifications to accelerate the optimization of implicit neural representations (INRs) for sparse-view CT reconstruction: (1) a filtered least squares (FLS) loss that applies a ramp-filter preconditioning to the data fidelity term, and (2) an ADMM algorithm that splits the problem into a linear projection-domain subproblem and a pixel-space INR fitting subproblem. The authors analyze the conditioning of the linearized least-squares problem with a fixed INR feature matrix, reporting that FLS reduces the Gram matrix condition number by 1-3 orders of magnitude at initialization for ReLU, SIREN, and hash-encoding architectures. In experiments on one simulated sparse-view breast CT phantom with 128 views and Poisson noise, they show that FLS and ADMM substantially reduce image MSE over 1000 iterations compared with standard least-squares training, achieving reconstructions visually similar to a tuned TV-regularized baseline.
Significance. If the reported acceleration holds under a fair cost comparison, the paper makes a useful contribution: FLS is a simple, architecture-agnostic preconditioning idea that can be plugged into any INR-based CT reconstruction, and the ADMM formulation decouples the expensive projection operator from the INR training, which could enable reuse of fast INR fitting machinery. The linearized conditioning analysis is the strongest element, being theoretically motivated and empirically verified over 100 random initializations. The experimental protocol is also careful in avoiding the inverse crime by simulating projections on a 2048x2048 grid while reconstructing on 512x512, and the comparison against a tuned TV baseline provides a meaningful reference. The paper is clearly written and the three INR architectures give some breadth to the empirical claims.
major comments (4)
- [Section 4, Fig. 2 and footnote] The computational cost comparison is not balanced for ADMM. For LS/FLS, one counted iteration equals one full gradient step, i.e., one application of P and P^T plus one Adam update. For ADMM, one counted iteration is a single CGLS step of the x-subproblem; each outer ADMM iteration additionally performs 50 Adam updates of the INR, whose forward/backward cost through the rasterized network is not represented on the x-axis. Moreover, the figure caption does not state whether the plotted MSE is the current x_k from the x-update or the INR output q_k = E{f_{theta_k}}, which matters because theta is updated only after each block of 50 CGLS steps. The claim that ADMM gives the lowest final MSE after 1000 iterations should be re-evaluated at equal wall-clock time or equal total FLOPs; please provide a timing comparison and clarify exactly which quantity is plotted.
- [Section 4, Fig. 2] All reconstruction MSE curves are obtained on a single synthetic breast CT phantom and a single noise realization, with no error bars or multiple seeds reported. The conditioning ratios in Table 1 are averaged over 100 initializations, but the central acceleration claim rests on the MSE curves, which are not. To support the conclusion that both approaches 'significantly accelerate' INR-based CT reconstruction, please report results over several noise realizations or phantoms, with statistical summaries (e.g., mean and standard deviation of final MSE or of MSE vs iteration).
- [Section 3.1, Eq. (4)] The ramp filter F used in the FLS loss has a zero at DC, so the quadratic form in Eq. (4) is a seminorm: constant offsets in z-y are not penalized. The manuscript does not discuss this nullspace or its effect on the linearized condition number analysis in Section 3.1.1 and Table 1. Please analyze this issue explicitly; either modify F to include a small DC penalty, or provide a numerical demonstration that the nullspace does not bias reconstructions in the reported setting.
- [Section 3.2] The ADMM updates for the nonconvex constraint x = E{f_theta} are not guaranteed to converge; the paper cites residual monitoring from [11] but reports no primal/dual residuals or stopping criterion. Since the experiments fix 20 outer iterations and 50 inner Adam/CGLS steps, it is unclear whether the 1000-iteration results correspond to convergence of the algorithm or to an arbitrary cutoff. Please report convergence diagnostics or otherwise justify the stopping rule.
minor comments (4)
- [Title] The title contains unwanted spacing ('ACCELERA TED OPTIMIZA TION', 'REPRESENTA TIONS'); please fix the typesetting.
- [Section 3.1.1, Eq. (6)] The expression f_theta(r) = PW i=1 ... appears to have a rendering artifact (the summation sign is missing); please ensure the equation is typeset correctly.
- [Section 4, hyperparameters] The learning rates and ADMM penalty parameters are tuned separately per architecture, and the results may be sensitive to these choices; please add a brief sensitivity discussion or ablation.
- [Figure 2] Since the y-axis is log-scale, it would be helpful to mark the outer ADMM iterations or to use distinct symbols for ADMM so that the reader can see where the INR subproblem is updated.
Circularity Check
No significant circularity: the acceleration claims are empirical comparisons against ground truth and baselines, not derived from the method's own assumptions.
full rationale
The paper's central claims are that a filtered least-squares loss and an ADMM-based optimizer accelerate INR-based sparse-view CT reconstruction. These are supported by measured reconstruction MSE relative to a ground-truth phantom, not by a chain of definitions that encode the outcome. The FLS loss is a proposed preconditioning objective; its condition-number analysis in Section 3.1.1 is a motivational rationalization, and Table 1 reports an empirical condition-number ratio rather than a fitted prediction. The ADMM algorithm in Section 3.2 is a standard application of a known optimization framework, with convergence monitored rather than assumed. Hyperparameters are tuned, but tuning does not make the comparison circular because the central evidence is the final MSE against an external ground truth and comparison with TV/FBP baselines. The only notable concern, flagged in the paper's own footnote, is that the iteration count for ADMM counts one CGLS step (one P and P^T product) while the INR update uses 50 Adam steps per outer iteration, so wall-clock comparisons could differ from iteration-count comparisons; this is a benchmarking-fairness issue, not a circularity in the derivation. The paper is self-contained against empirical baselines and does not reduce its predictions to its inputs.
Assumptions & free parameters
free parameters (5)
- ADMM penalty parameter μ =
ReLU: 1, SIREN: 3, Hash Encoding: 2.5
- Optimizer learning rates =
ReLU INR: 1e-3; SIREN/Hash: 1e-4; ADMM inner Adam: 1e-4 (ReLU/SIREN), 1e-3 (Hash)
- SIREN base frequency ω0 =
75
- Fourier feature max frequency =
‖k‖ ≤ 15, frequencies 2πk
- ADMM inner iteration counts =
20 outer, 50 Adam, 50 CGLS
assumptions (4)
- domain assumption The rasterized fan-beam projection matrix P accurately models the CT acquisition
- standard math The ramp filter matrix F is positive semi-definite
- domain assumption The linearized analysis with fixed Q and only final-layer weights optimized is representative of full INR training
- domain assumption ADMM with inexact solves and a nonlinear constraint is practically convergent when residuals are monitored
Cite this review
Pith. "Pith review of Accelerated Optimization of Implicit Neural Representations for CT Reconstruction." pith.science (2026). https://pith.science/paper/OPNGSAWJ
@misc{pith2026250413390,
author = {Pith},
title = {Pith review of: Accelerated Optimization of Implicit Neural Representations for CT Reconstruction},
year = {2026},
howpublished = {\url{https://pith.science/paper/OPNGSAWJ}},
note = {Machine review of arXiv:2504.13390}
}
read the original abstract
Inspired by their success in solving challenging inverse problems in computer vision, implicit neural representations (INRs) have been recently proposed for reconstruction in low-dose/sparse-view X-ray computed tomography (CT). An INR represents a CT image as a small-scale neural network that takes spatial coordinates as inputs and outputs attenuation values. Fitting an INR to sinogram data is similar to classical model-based iterative reconstruction methods. However, training INRs with losses and gradient-based algorithms can be prohibitively slow, taking many thousands of iterations to converge. This paper investigates strategies to accelerate the optimization of INRs for CT reconstruction. In particular, we propose two approaches: (1) using a modified loss function with improved conditioning, and (2) an algorithm based on the alternating direction method of multipliers. We illustrate that both of these approaches significantly accelerate INR-based reconstruction of a synthetic breast CT phantom in a sparse-view setting.
Reference graph
Works this paper leans on
-
[11]
Wire: Wavelet implicit neural representations,
Vishwanath Saragadam, Daniel LeJeune, Jasper Tan, Guha Balakrishnan, Ashok Veeraraghavan, and Richard G Baraniuk, “Wire: Wavelet implicit neural representations,” in Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition , 2023, pp. 18507–18516
work page 2023
-
[1]
INTRODUCTION Machine learning (ML) has shown promise in solving many challenging ill-posed image reconstruction problems in X-ray computed tomography (CT). Much of the attention has been given to supervised or semi-supervised approaches, which require access to “ground truth” target images. However, in many imaging applications, there is limited access to...
-
[2]
Accelerated Optimization of Implicit Neural Representations for CT Reconstruction
PROBLEM FORMULA TION Suppose we are given sinogram measurements y∈ Rm (re- arranged into a vector). Letfθ : Rd→ R denote a given INR architecture with trainable parametersθ∈ Rp. The goal is to find a set of INR parameters θ∈ Rp such thatP{fθ}≈ y, whereP is a linear operator approximating line integrals of the function fθ over a set of lines dictated by th...
work page Pith review arXiv 2025
-
[3]
PROPOSED APPROACHES 3.1. Filtered Least Squares Loss Function To address the ill-conditioning arising when using a least squares loss, we propose using a modified least squares loss that incorporates an “FBP-type” preconditioning matrix. We note that such preconditiong strategies have a long history in MBIR for CT reconstruction; see [6] and references th...
-
[4]
EXPERIMENTS We illustrate the proposed INR training modifications for the reconstruction of a breast CT phantom from simulated noisy sparse-view data. We use a modified version of the phantom introduced in [12]; the phantom is modified to have texture, and a smooth deformation is applied such that the boundary of the breast is non-circular. Sinogram data ...
-
[5]
DISCUSSION AND CONCLUSION This work proposed using a modified loss function (filtered least squares) and a modified training algorithm (nonlinear constrained ADMM) for efficient training of INRs in CT re- construction. Our results on simulated breast CT data show that both approaches substantially accelerate INR optimiza- tion relative to directly minimiz...
-
[6]
COMPLIANCE WITH ETHICAL STANDARDS This is a numerical simulation study for which no ethical ap- proval was required
-
[7]
ACKNOWLEDGMENTS This work was supported by NSF CRII award CCF-2153371
Show all 25 references
-
[8]
CoIL: Coordinate-based in- ternal learning for tomographic imaging,
Yu Sun, Jiaming Liu, Mingyang Xie, Brendt Wohlberg, and Ulugbek S Kamilov, “CoIL: Coordinate-based in- ternal learning for tomographic imaging,” IEEE Trans- actions on Computational Imaging , vol. 7, pp. 1400– 1412, 2021
2021
-
[9]
NeRP: implicit neural representation learning with prior embedding for sparsely sampled image reconstruction,
Liyue Shen, John Pauly, and Lei Xing, “NeRP: implicit neural representation learning with prior embedding for sparsely sampled image reconstruction,” IEEE Trans- actions on Neural Networks and Learning Systems, vol. 35, no. 1, pp. 770–782, 2022
2022
-
[10]
PINER: Prior- informed implicit neural representation learning for test- time adaptation in sparse-view CT reconstruction,
Bowen Song, Liyue Shen, and Lei Xing, “PINER: Prior- informed implicit neural representation learning for test- time adaptation in sparse-view CT reconstruction,” in Proceedings of the IEEE/CVF winter conference on ap- plications of computer vision, 2023, pp. 1928–1938
2023
-
[12]
An attenuation field network for dedicated cone beam breast CT with short scan and offset detector ge- ometry,
Zhiyang Fu, Hsin Wu Tseng, and Srinivasan Vedan- tham, “An attenuation field network for dedicated cone beam breast CT with short scan and offset detector ge- ometry,”Scientific Reports, vol. 14, no. 1, pp. 319, 2024
2024
-
[13]
A fast regularized iterative algorithm for fan-beam CT reconstruction,
Ting Wang, Hiroyuki Kudo, Fukashi Yamazaki, and Huafeng Liu, “A fast regularized iterative algorithm for fan-beam CT reconstruction,” Physics in Medicine & Biology, vol. 64, no. 14, pp. 145006, 2019
2019
-
[14]
Amir Beck, Introduction to nonlinear optimization: Theory, algorithms, and applications with MATLAB , SIAM, 2014
2014
-
[15]
ReLUs are sufficient for learning implicit neural repre- sentations,
Joseph Shenouda, Yamin Zhou, and Robert D Nowak, “ReLUs are sufficient for learning implicit neural repre- sentations,” arXiv preprint arXiv:2406.02529, 2024
2024 arXiv
-
[16]
Distributed optimization and statistical learning via the alternating direction method of multipliers,
Stephen Boyd, Neal Parikh, Eric Chu, Borja Peleato, Jonathan Eckstein, et al., “Distributed optimization and statistical learning via the alternating direction method of multipliers,” Foundations and Trends® in Machine learning, vol. 3, no. 1, pp. 1–122, 2011
2011
-
[17]
LSQR: An algorithm for sparse linear equations and sparse least squares,
Christopher C Paige and Michael A Saunders, “LSQR: An algorithm for sparse linear equations and sparse least squares,” ACM Transactions on Mathematical Software (TOMS), vol. 8, no. 1, pp. 43–71, 1982
1982
-
[18]
Nonconvex gener- alization of alternating direction method of multipliers for nonlinear equality constrained problems,
Junxiang Wang and Liang Zhao, “Nonconvex gener- alization of alternating direction method of multipliers for nonlinear equality constrained problems,” arXiv preprint arXiv:1705.03412, 2017
2017 arXiv
-
[19]
Do CNNs solve the CT inverse prob- lem?,
Emil Y . Sidky, Iris Lorente, Jovan G. Brankov, and Xi- aochuan Pan, “Do CNNs solve the CT inverse prob- lem?,” IEEE Transactions on Biomedical Engineering, vol. 68, no. 6, pp. 1799–1810, 2021
2021
-
[20]
Convex optimization problem prototyping for image reconstruction in computed tomography with the chambolle–pock algorithm,
Emil Y Sidky, Jakob H Jørgensen, and Xiaochuan Pan, “Convex optimization problem prototyping for image reconstruction in computed tomography with the chambolle–pock algorithm,” Physics in Medicine & Bi- ology, vol. 57, no. 10, pp. 3065, 2012
2012
-
[21]
Fourier features let networks learn high fre- quency functions in low dimensional domains,
Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Sing- hal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng, “Fourier features let networks learn high fre- quency functions in low dimensional domains,” Ad- vances in neural informat...
2020
-
[22]
Implicit neural representations with periodic activation functions,
Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein, “Implicit neural representations with periodic activation functions,” Ad- vances in Neural Information Processing Systems , vol. 33, pp. 7462–7473, 2020
2020
-
[23]
Instant neural graphics primitives with a multiresolution hash encoding,
Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexander Keller, “Instant neural graphics primitives with a multiresolution hash encoding,” ACM transac- tions on graphics (TOG), vol. 41, no. 4, pp. 1–15, 2022
2022
-
[24]
Accelerated image reconstruction using ordered subsets of projection data,
H Malcolm Hudson and Richard S Larkin, “Accelerated image reconstruction using ordered subsets of projection data,” IEEE Transactions on Medical Imaging, vol. 13, no. 4, pp. 601–609, 1994
1994
-
[25]
Learned initializations for optimizing coordinate-based neural representations,
Matthew Tancik, Ben Mildenhall, Terrance Wang, Divi Schmidt, Pratul P Srinivasan, Jonathan T Barron, and Ren Ng, “Learned initializations for optimizing coordinate-based neural representations,” in Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni...
2021
Reviewed August 16, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.