{"id":"7902b26d-b32f-4c2f-ad07-78374cab717b","arxiv_id":"2501.11586","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"PCA compression of the redundancy weight layer reduces trainable parameters by 97.25% in a differentiable shift-variant FBP model with similar reconstruction quality on simulated sinusoidal-trajectory CBCT.","lead":"This paper compresses the learned redundancy weights in a differentiable filtered backprojection network for CT reconstruction, cutting trainable parameters by 97.25% using PCA. The compressed model reconstructs simulated cone-beam CT images about as well as the original on a sinusoidal C-arm trajectory.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Training-speed claim ignores the one-time cost of training the uncompressed model to obtain the PCA basis; the 430-to-140 epoch comparison is not end-to-end.","rationale":"The reader's verdict is CONDITIONAL with MODERATE confidence, and the rationale already notes that the training speed claim uses epochs rather than wall-clock time. My concern sharpens this: the epoch comparison is not merely missing wall-clock time; it likely omits the dominating pretraining cost needed to obtain the PCA basis. This reinforces the need for a conditional verdict rather than acceptance. I do not see a fatal flaw in the parameter-reduction claim itself: Table 1's parameter counts are consistent with V_k and mu being trainable and counted, and the empirical accuracy on the tested sinusoidal trajectory is promising, though the small validation set (6 samples) and the SSIM drop temper the 'without compromising accuracy' claim. The PCA-subspace concern raised by the reader is partially mitigated if V_k and mu are trainable, since the model can adapt the subspace during training; the unresolved issue is whether the low-rank constraint remains sufficient when the model is trained from scratch or on other trajectories. The most concrete, load-bearing gap is the end-to-end training time comparison, which is directly testable from the authors' own pipeline. My recommendation is UNCHANGED because the verdict should remain CONDITIONAL, pending the requested wall-clock and pretraining-cost accounting.","tokens_in":4368,"tokens_out":13033,"duration_ms":136142,"concrete_test":"Measure wall-clock time for (a) training the uncompressed model to convergence, (b) computing PCA from its redundancy weights, and (c) training the compressed model with PCA initialization to convergence. Compare (a) against (b)+(c). Also report per-epoch times and the convergence time for the compressed model initialized without PCA (e.g., random V_k and mu) to separate initialization benefit from low-rank regularization. If (b)+(c) exceeds (a), the training-speed claim fails unless the PCA basis can be reused across trajectories or computed from a cheap proxy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central practical claim is that the compressed model 'greatly improves training speed.' Section 4 reports convergence in 140 epochs versus 430 for the uncompressed network. However, the PCA basis V_k and mean mu in Eq. (7) are computed in Section 2.3 from redundancy weights 'learned from sinusoidal trajectory projection data'—i.e., from an already-trained uncompressed model. The compressed model is then initialized with this PCA decomposition and trained (Figure 2). Thus the total training cost of the proposed pipeline is at least the cost of the uncompressed training plus the compressed training, not less than the uncompressed training alone. The paper only compares epochs, not wall-clock time, and does not report per-epoch times or the pretraining cost. If the PCA basis cannot be obtained cheaply or amortized across trajectories, the claimed speedup is unsupported. This is load-bearing because the abstract and conclusion advertise faster training as a main benefit. Secondary: Table 1 shows SSIM dropping from 0.9591 to 0.9481 on only 6 validation samples with no significance test, so 'without compromising reconstruction accuracy' is also not rigorously established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a PCA-based compression scheme for the redundancy-weight layer of a differentiable shift-variant filtered backprojection (FBP) reconstruction network. The authors train an uncompressed model on simulated cone-beam CT data acquired along a sinusoidal trajectory, compute a PCA basis from the trained redundancy weights, and then replace the high-dimensional redundancy weight layer with a low-dimensional trainable coefficient vector mapped back through the fixed PCA basis (Eq. 7). They report a 97.25% reduction in trainable parameters and claim that reconstruction accuracy is preserved and training is greatly accelerated (140 epochs vs. 430 epochs). Experimental results are presented for one sinusoidal trajectory with 24 training and 6 validation samples.","tokens_in":4578,"tokens_out":4406,"duration_ms":47755,"significance":"If the central claims hold, the work is a practical contribution: it offers a simple, operator-level compression for a reconstruction network that otherwise stores a large redundancy weight tensor, and it suggests that this layer has low intrinsic dimensionality. The authors provide quantitative metrics in Table 1, visual comparisons in Figures 3-4, and a clear parameter-count reduction. However, the evidence currently does not fully support the two headline claims: the training-time comparison is not end-to-end, because obtaining the PCA basis requires a fully trained uncompressed model, and the accuracy comparison shows a small but consistent SSIM drop with no statistical assessment. The generalizability to other trajectories is also untested, which limits the practical claim. These issues are addressable and do not invalidate the core idea, but they need to be fixed before publication.","major_comments":[{"comment":"The claim that the method 'significantly accelerates the training process' is not supported end-to-end. The PCA basis V_k and mean mu are computed from the redundancy weights of an already-trained uncompressed model (Section 2.3), so the total cost of the proposed pipeline is at least the cost of training the uncompressed model plus training the compressed model. Comparing 140 epochs versus 430 epochs only measures the compressed training phase, not the whole procedure. Please report wall-clock training times, including the PCA estimation and any overhead, or clearly state the amortization scenario under which the speedup is realized (e.g., reusing the basis across many trajectories).","section":"Section 4 and Section 2.3, Eq. (7)"},{"comment":"The statement 'without compromising reconstruction accuracy' is not rigorously established. SSIM drops from 0.9591 +/- 0.0051 (No PCA) to 0.9427 +/- 0.0088 for PCA(50) and 0.9481 +/- 0.0074 for PCA(10), while MSE and PSNR changes are small and in the opposite direction. With only six validation samples and no paired significance test, it is unclear whether these differences are statistically meaningful or reflect a tradeoff among metrics. Please report per-sample results, confidence intervals, or a paired test, and temper the accuracy claim accordingly.","section":"Table 1"},{"comment":"It is ambiguous whether V_k is trainable. The abstract says 'trainable eigenvector matrix,' but Eq. (7) shows only w'_red as the low-dimensional variable and treats V_k and mu as fixed quantities from PCA. This distinction is essential for the parameter-count claim: if V_k is frozen, the reported 3,116,560 parameters for PCA(10) make sense, but if V_k is trained, its entries must be included in the trainable parameter count and the 97.25% reduction may change. Please state explicitly which quantities are frozen and which are trained, and reconcile the parameter counts with that statement.","section":"Abstract and Section 2.3, Eq. (7)"},{"comment":"The generalizability claim is not supported by the experimental design. The PCA basis is estimated from the redundancy weights learned on a single sinusoidal trajectory, and the compressed model is tested on that same trajectory only. Since the paper motivates the method for non-circular and arbitrary trajectories, at least one additional trajectory (or a clear statement that cross-trajectory transfer is out of scope) is needed. The current 24-sample training set and 6-sample validation set also make it difficult to draw strong conclusions about reconstruction quality.","section":"Section 3"}],"minor_comments":[{"comment":"The text describes standardization, but Eq. (3) only subtracts the mean and does not divide by the standard deviation. Please align the description with the equations or clarify that no scaling is applied.","section":"Section 2.3, Eqs. (3)-(5)"},{"comment":"The symbols w_cos, w_sino, w_red, and w_d are not defined. Please define each weight matrix and indicate its role in the reconstruction pipeline relative to Eq. (1).","section":"Section 2.2, Eq. (2)"},{"comment":"The integration domain 'S^2/2' is unclear. Please specify the half-sphere notation or give a reference for this representation.","section":"Section 2.1, Eq. (1)"},{"comment":"The paper does not report the parameters of the sinusoidal trajectory (e.g., amplitude, frequency, number of projections, detector geometry) or the forward projector settings. This information is necessary for reproducibility.","section":"Section 3"},{"comment":"The sentence 'while the remaining samples were employed for validation' leaves the validation count implicit. Please state that only 6 samples were used for validation.","section":"Section 3"},{"comment":"Figure 1 is referenced in Section 2.3 but no visible figure is provided in the text; please ensure the figure is included and that its axis labels and legend are readable.","section":"Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is built on the authors' own prior work (refs. 3 and 4), and the baseline is not independently reproduced. This is acceptable, but given that the baseline is also the source of the PCA basis, the reader should be aware that the compression analysis is inherently tied to that specific model. The paper appears to be written in a short conference format; the missing experimental details (trajectory parameters, validation count, wall-clock runtime) are likely within the authors' reach to supply in a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a niche but legitimate compression result for a specialized CT reconstruction network. The 97% parameter reduction is real, but the two headline claims—no accuracy loss and much faster training—are both overstated. The speed claim is the bigger problem: the PCA basis comes from an already-trained uncompressed model, so the end-to-end training cost is not less than the uncompressed baseline; you pay for the uncompressed training first. The paper compares epochs (140 vs 430) and never reports wall-clock time or the pretraining cost. That is a load-bearing flaw in the main practical selling point.\n\nWhat is genuinely new: applying PCA to the redundancy-weight layer of the differentiable shift-variant FBP, rather than to generic network weights. The paper shows that the learned redundancy weights live in a low-dimensional subspace—that is a real empirical finding on a non-trivial model, and Figure 1 is a useful diagnostic. The math is simple and correctly stated, and the paper does not hide the fact that quality degrades with smaller k.\n\nThe soft spots are mostly presentation and scope. Table 1 shows SSIM dropping from 0.959 to 0.943/0.948, so \"without compromising accuracy\" is too strong. The validation set is six volumes, no significance test, and only one sinusoidal trajectory is tested. The paper does not cite or compare against standard network-compression literature (pruning, low-rank factorization, distillation), so the novelty is genuinely incremental. There is also a technical ambiguity: the abstract calls V_k a \"trainable eigenvector matrix,\" but Eq. 7 and the parameter counts suggest V_k is either fixed or shared; if it is trainable as a full matrix, the parameter savings do not add up. The reader should check that.\n\nWho should read it: researchers working on learned non-circular trajectory CBCT reconstruction. For them, the compressibility observation and the PCA reparametrization are useful. For the broader ML community, this is a small case study.\n\nRecommendation: I would send it to peer review, but with a clear request for major revision: reframe the speed claim, add wall-clock and end-to-end cost, include at least one standard compression baseline, fix the V_k ambiguity, and report significance or at least more validation samples. The core observation is sound enough to deserve referee time.","headline":"Niche but plausible compression result for a specialized CT model; the speed and accuracy claims overreach, especially the ignored pretraining cost.","tokens_in":5132,"tokens_out":3238,"would_cite":false,"duration_ms":34202,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The redundancy weights of differentiable shift-variant FBP can be replaced by a PCA decomposition, cutting trainable parameters 97.25% with no loss in reconstruction quality.","keywords":["CT reconstruction","network compression","deep learning","known operator","PCA","arbitrary trajectory","shift-variant FBP","cone-beam computed tomography"],"falsifier":"Train the compressed model on a different non-circular trajectory, such as an elliptical or saddle orbit, using a PCA basis estimated on that trajectory, and compare reconstruction quality with the uncompressed model. If MSE, PSNR, or SSIM degrades substantially more than in Table 1, the claim that redundancy weights are generally low-rank is falsified for that setting; re-running the same sinusoidal setup from different random initializations would also test how much of the 97.25% figure depends on the specific trained model used to estimate the PCA basis.","tokens_in":4118,"feed_emoji":"🩻","tokens_out":6431,"duration_ms":66577,"temperature":0.7,"pith_summary":"This paper sets out to show that the redundancy-weight layer of a differentiable shift-variant filtered backprojection (FBP) reconstruction network is highly compressible, and that replacing it with a principal-component decomposition preserves reconstruction quality. The authors integrate PCA directly into the pipeline, representing the redundancy weights as a low-dimensional coefficient vector mapped back through a principal-component basis plus the mean (Eq. 7). On a sinusoidal C-arm trajectory with 24 training samples, the compressed model cuts trainable parameters from 113,184,000 to 3,116,560 (a 97.25% reduction) while keeping MSE, PSNR, and SSIM comparable to the uncompressed model, and converges in about 140 epochs instead of 430. If the result holds beyond this single trajectory, it would make non-circular cone-beam CT reconstruction practical on resource-constrained hardware.","feed_headline":"PCA cuts CT reconstruction training weights by 97.25%","feed_subtitle":"Redundancy-weight layer becomes a 10-component linear map; quality holds while epochs drop from 430 to 140.","key_machinery":"The central object is the PCA decomposition of the redundancy-weight layer. Starting from the principal components $V_k$ and mean $\\mu$ of the redundancy weights of an already-trained model, the pipeline replaces the high-dimensional weight layer with a trainable low-dimensional code $w'_{red}$ and the linear mapping $w'_{red} V_k^T + \\mu$, inserted at the position of $w_{red}$ in the differentiable FBP chain (Eq. 7). The number of principal components $k$ controls the trade-off: $k = 50$, $30$, and $10$ yield 14,450,960, 8,783,760, and 3,116,560 trainable parameters respectively, with negligible changes in image quality.","core_discovery":"The central claim is that the redundancy weights in the differentiable shift-variant FBP model live on a low-dimensional linear subspace, so the learned weight tensor can be replaced by the reconstruction $w'_{red} V_k^T + \\mu$ without sacrificing reconstruction accuracy. With only ten principal components the method retains image quality relative to the full model (MSE 0.0855 vs 0.0904, PSNR 37.67 dB vs 37.20 dB, SSIM 0.9481 vs 0.9591), while the number of trainable parameters falls from 113,184,000 to 3,116,560. The paper further reports that the compressed network reaches convergence in 140 epochs, compared with 430 for the uncompressed network, and that the redundancy weights recovered from the low-dimensional representation closely match the weights learned without compression.","pith_inferences":["A natural extension the paper leaves implicit is cross-trajectory transfer: a PCA basis estimated from one orbit might serve as an initialization for another orbit, which would amortize the cost of computing the basis; this is testable by fitting $V_k$ on a sinusoidal trajectory and evaluating on an elliptical one.","Because PCA is linear, the success of this compression suggests that a spectral analysis of learned reconstruction weights could predict compressibility before full training, potentially selecting $k$ automatically from the eigenvalue decay.","The same low-rank substitution could be applied to the other learned weighting layers in the pipeline ($w_{cos}$, $w_{sino}$, $w_d$), which may yield additional parameter reductions beyond the 97.25% reported for the redundancy weights alone.","The evaluation on a single sinusoidal trajectory with 24 training samples means the numerical claims are best read as a proof of concept; testing on more trajectories and more samples would establish whether 97.25% is a typical compression rate or a favorable one."],"forward_implications":["With PCA(10), the differentiable shift-variant FBP model trains to convergence in roughly one-third of the epochs (140 vs 430), directly lowering the compute cost of learning redundancy weights for a given trajectory.","The parameter reduction from 113,184,000 to 3,116,560 makes the reconstruction network small enough for memory-constrained deployment, where the uncompressed model would be impractical.","Reconstruction quality stays within the same range as the uncompressed model (SSIM 0.9481 vs 0.9591), so the compressed pipeline can be substituted without a visible quality trade-off on this trajectory.","The same PCA insertion applies to any trajectory for which a redundancy-weight model has been trained, because the compression acts on the learned weights rather than on the analytic formula.","Larger $k$ values (50, 30) give a safety margin: even a 30-component model cuts parameters from 113,184,000 to 8,783,760 with SSIM 0.9409, leaving room to trade size against fidelity."],"supporting_citations":[{"why":"Supplies the shift-variant FBP reconstruction formula whose redundancy weights are the target of compression.","marker":"[2]"},{"why":"Establishes the differentiable shift-variant FBP model that learns redundancy weights via backpropagation, the base architecture this paper compresses.","marker":"[3]"},{"why":"Extends the differentiable shift-variant FBP model to arbitrary CBCT orbits, the setting for the compressed pipeline.","marker":"[4]"},{"why":"Provides the PCA theory used to reduce and reconstruct the redundancy-weight parameters.","marker":"[5]"},{"why":"Supplies the cone-beam forward projection used to generate the simulated training and validation data.","marker":"[7]"}],"fun_headline_variants":["PCA cuts CT model parameters 97.25%, quality holds","97% fewer CT training weights via PCA, same reconstruction quality","PCA reduces CT reconstruction weights 97%, training epochs 430→140","Ten PCA components replace 113M CT weights, same accuracy","PCA compresses shift-variant FBP weights 97.25% without quality loss"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the PCA basis and mean computed from the redundancy weights of an already-trained uncompressed model remain a sufficient coordinate system for the redundancy weights during training of the compressed model, so the network does not need to leave that low-dimensional subspace; this is tested on only one sinusoidal trajectory with 24 training samples.","fun_headline_variants_meta":{"raw":{"variants":["PCA cuts CT model parameters 97.25%, quality holds","97% fewer CT training weights via PCA, same reconstruction quality","PCA reduces CT reconstruction weights 97%, training epochs 430→140","Ten PCA components replace 113M CT weights, same accuracy","PCA compresses shift-variant FBP weights 97.25% without quality loss"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001495,"raw_usage":{"total_tokens":5989,"prompt_tokens":920,"completion_tokens":5069,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":4976}},"tokens_in":536,"tokens_out":5069,"duration_ms":37462,"temperature":1.0,"reasoning_tokens":4976,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:04:02.287740+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the compressed model on a different non-circular trajectory, such as an elliptical or saddle orbit, using a PCA basis estimated on that trajectory, and compare reconstruction quality with the uncompressed model. If MSE, PSNR, or SSIM degrades substantially more than in Table 1, the claim that redundancy weights are generally low-rank is falsified for that setting; re-running the same sinusoidal setup from different random initializations would also test how much of the 97.25% figure depends on the specific trained model used to estimate the PCA basis.","supporting_citations":[{"cited_title":"and Clack, R., ``A cone-beam reconstruction algorithm using shift-variant filtering and cone-beam backprojection,'' IEEE transactions on medical imaging 13 (1), 186--195 (1994)","cited_arxiv_id":null,"evidence_quote":"Supplies the shift-variant FBP reconstruction formula whose redundancy weights are the target of compression."},{"cited_title":"Deep Learning Computed Tomography based on the Defrise and Clack Algorithm","cited_arxiv_id":"2403.00426","evidence_quote":"Establishes the differentiable shift-variant FBP model that learns redundancy weights via backpropagation, the base architecture this paper compresses."},{"cited_title":"DRACO: Differentiable Reconstruction for Arbitrary CBCT Orbits","cited_arxiv_id":"2410.14900","evidence_quote":"Extends the differentiable shift-variant FBP model to arbitrary CBCT orbits, the setting for the compressed pipeline."},{"cited_title":"and Ratajczak, W., ``Principal components analysis (pca),'' Computers & Geosciences 19 (3), 303--342 (1993)","cited_arxiv_id":null,"evidence_quote":"Provides the PCA theory used to reduce and reconstruct the redundancy-weight parameters."},{"cited_title":"K., ``Pyro-nn: Python reconstruction operators in neural networks,'' Medical physics 46 (11), 5110--5115 (2019)","cited_arxiv_id":null,"evidence_quote":"Supplies the cone-beam forward projection used to generate the simulated training and validation data."}],"review_version":1}