{"id":"98f3e943-b9bd-44e5-9954-b5f9daa25f9d","arxiv_id":"2505.08324","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"incDG inserts a deep-network estimate at every incremental TpV step, giving faster and more stable deblurring and CT reconstructions than either the pure model-based or pure deep-learning baselines, with slightly lower CT accuracy than the model-based solver.","lead":"incDG is a hybrid image-reconstruction algorithm that alternates a neural network's quick estimate with a small number of steps of a classical optimization solver inside an incremental sparsity schedule. In tests on deblurring and sparse-view CT, it is faster and noticeably more stable than the pure network baseline, though the pure model-based solver still gives the best CT pixel accuracy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim of ell0-optimality rests on an inexact incremental scheme whose convergence Section 2.4 explicitly disclaims; incDG's 5-CP-iteration stages may be a tuned network cascade rather than a convergent solver.","rationale":"The reader's conditional verdict already identifies the same load-bearing assumption: the inexact incremental schedule with unproven convergence is what carries the claim that incDG approximates the ell0-optimal solution. My reading confirms this and adds a concrete mechanism: with kCP = 5 and one IR step per stage, the inner subproblems are not solved to any controlled accuracy, so the method's behavior is dominated by the network initialization and the hand-set schedule. The paper itself flags the missing convergence proof in Section 2.4, so this is not an external standard being imposed; it is an internal admission that the theory does not cover the implemented algorithm. The empirical overclaim in the CT section is also real, but it is a claim-reframing issue rather than the deepest vulnerability. I would keep the reader's CONDITIONAL verdict: the method may well be useful as a heuristic accelerator, but the abstract's ell0-optimality and theoretical-guarantee statements need either proof or explicit softening. No stronger action is warranted because the deblurring experiments and the stability advantage over incNN are credible empirical evidence of practical value.","tokens_in":19213,"tokens_out":3855,"duration_ms":40122,"concrete_test":"Run the incDG algorithm on the same COULE and Mayo test sets with the published configuration and with progressively more exact inner solves, e.g., K = [5,5,5,5] versus [50,50,50,50] versus [500,500,500,500], or solve each subproblem (11) to high precision with the stopping criteria tau_x = tau_F = 1e-7, keeping all trained networks and other hyperparameters fixed. If final RE/SSIM and the TpV objective (15) are statistically unchanged across these schedules, the MB refinement is not responsible for incDG's results and the ell0-optimality claim is unsupported. If the objective decreases monotonically and metrics improve as inner iterations increase, the optimization component is load-bearing and the concern is weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.4 states that inexactly solving each outer step by simultaneously updating lambda(h) and p(h) has no proven convergence, and that in 'our setting' each TpV problem is stopped after only one iteration. Algorithm 3 then embeds these truncated subproblems behind a network: with K = [5,5,5,5] (deblurring) or [5,5,5,5,5,5] (CT), each stage performs one IR step with kCP = 5 Chambolle-Pock iterations. No bound quantifies the error from this truncation, and no result shows the final iterate is stationary for problem (15) or a local minimizer of the TpV functional. Consequently the abstract's 'efficiently approximates the ell0-optimal solution' and 'theoretical guarantees of model-based optimization' are unsupported: the output could be dominated by the pretrained networks and the arbitrary choice of alpha_p, lambda(0), and scheduler K, rather than by convergence of the incremental optimization. The paper's own CT results (Figures 8-9) reinforce this: incTpV achieves the best RE/SSIM, so incDG does not 'outperform' conventional iterative solvers in accuracy there; its advantage is speed and stability relative to incNN. The central claim therefore needs either a convergence or suboptimality analysis for the inexact schedule, or a reframing to 'heuristic acceleration' with empirical claims restricted to speed and stability.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes incDG, a hybrid algorithm for non-convex TpV-regularized inverse problems in medical imaging. It combines an incremental reweighted-ell_1 scheme with per-stage ResUNet 'Deep Guess' initializations followed by a few Chambolle-Pock iterations. The stated claims are that incDG efficiently approximates the ell_0-optimal solution, outperforms both conventional iterative solvers and deep-learning-based methods in accuracy and stability, and can be trained without ground truth. Experiments cover deblurring on COULE images with Brain CT generalization tests, and sparse-view CT reconstruction on the Mayo Clinic dataset.","tokens_in":19498,"tokens_out":5046,"duration_ms":48275,"significance":"If the main claims were fully established, incDG would be a practically valuable hybrid: it mitigates the instability of pure network outputs by adding a small number of model-based iterations, and the ground-truth-free training variant would ease clinical deployment. The paper has real strengths: it uses external public datasets, compares incNN, incTpV, and incDG under matched parameter settings, reports wall-clock timings, and provides a clear empirical demonstration that a few MB iterations stabilize a network-based reconstruction. However, the central theoretical claim is explicitly disclaimed in Section 2.4, and the CT results contradict the abstract's blanket accuracy claim. The paper needs either a convergence or suboptimality analysis for the inexact schedule, or a reframing of the contribution as heuristic acceleration with empirical claims restricted to speed and stability.","major_comments":[{"comment":"Section 2.4 states that 'the convergence of the resulting algorithm is not proven' and that in the authors' setting each TpV problem is stopped after only one iteration. Algorithm 3 then fixes kCP=5 and uses schedulers K=[5,5,5,5] for deblurring or [5,5,5,5,5,5] for CT. No error bound is given for this truncation, and no result shows that the final iterate is stationary for problem (15) or a local minimizer of the TpV functional. Consequently the abstract's 'efficiently approximates the ell_0-optimal solution' and Section 5's 'theoretical guarantees of model-based optimization' are unsupported. Please either provide a convergence or suboptimality analysis for the inexact schedule, or explicitly reframe incDG as a heuristic acceleration and limit the claims to the empirical results.","section":"§2.4, Algorithm 3"},{"comment":"The Abstract claims that 'incDG outperforms both conventional iterative solvers and deep learning-based methods, achieving superior accuracy and stability.' In the CT experiments, however, Figure 8's caption states that incTpV 'outperforms the fast methods, achieving the best RE and SSIM values on the displayed images,' and Figure 9's boxplots show incTpV with the best metric distributions over the test set. Thus the accuracy part of the claim is contradicted in the CT setting; incDG's advantage there is speed (Table 2) and stability relative to incNN, not accuracy relative to incTpV. Please restrict the accuracy claim to the deblurring experiments or revise the abstract and conclusions accordingly.","section":"Figure 8, Figure 9, Abstract"},{"comment":"The performance of incDG depends on several hand-set hyperparameters (lambda_0, alpha_p, H, scheduler K, kCP) that are fixed to 'achieve good performance on the training samples,' with no sensitivity study. Because Section 2.4 provides no convergence guarantee, the observed gap between incDG and incTpV/incNN could be a tuned artifact of the scheduler and network cascade rather than an intrinsic property of the algorithm. Please report a sensitivity analysis for at least alpha_p and the scheduler K, and quantify how the results change with kCP, to substantiate the claimed robustness.","section":"§4.1, Table 2"}],"minor_comments":[{"comment":"The stopping condition 'or count< kIR' is inverted: since count starts at 0, this condition is true immediately and would stop the loop at once. It should be 'or count>= kIR' (or an equivalent maximum-iteration check), with count accumulating kCP per outer iteration as described.","section":"Algorithm 1, line 10"},{"comment":"Equations (13) and (14) use '∀j' where the running index is h; moreover, Section 2.4 refers to 'Equation (5)' for the lambda and p updates, but the relevant formulation is Equation (15).","section":"Equations (13), (14), §2.4"},{"comment":"There are typographical slips such as 'tollerances' in Algorithms 2 and 3 and 'regularizaion' in Section 2.1 that should be corrected before publication.","section":"General typography"}],"recommendation":"major_revision","confidential_remarks":"The manuscript leans heavily on the authors' own prior work ([33], [35], [21]) for both the incremental scheme and the Deep Guess mechanism; this is not disqualifying, but the novelty claim would be stronger with a more detailed comparison against those baselines. The main issue for the editor is the mismatch between the paper's advertised theoretical guarantees and the explicit disclaimer in Section 2.4, combined with the CT results that contradict the abstract. These are fixable by reframing the claims, but they must be fixed before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Elena Morotti's paper is a decent engineering contribution that gets oversold in the abstract. The new thing is incDG: at each incremental TpV stage, a ResUNet produces a warm start, and then you run a handful of Chambolle-Pock iterations. This is a natural extension of her group's earlier incremental and Deep Guess work, and the experiments are the main value. On deblurring, incDG clearly beats both the purely model-based incTpV and the pure network cascade (incNN) in RE and SSIM, with much lower variance than incNN. The ground-truth-free training variant also looks practical: training the networks on incTpV solutions instead of clean images costs little. In CT, incDG gives an order-of-magnitude speedup over incTpV (4.5 s vs 53 s) and is far more stable than incNN. That's a useful result for clinical settings.\n\nThe soft spots are real. The abstract says incDG 'outperforms both conventional iterative solvers and deep learning-based methods' and 'achiev[es] superior accuracy and stability.' On the CT task, that is false: incTpV has the best RE/SSIM in the displayed images and in the boxplots. incDG's advantage there is speed, not accuracy. More importantly, the 'theoretical guarantees' language is not supported. Section 2.4 explicitly says the inexact incremental scheme has no proven convergence, and the implemented algorithm stops each TpV subproblem after one IR step with five CP iterations. No bound quantifies the truncation error, and no result shows the final iterate is stationary for the non-convex objective. So calling incDG an approximation to the ell0-optimal solution is a heuristic claim, not a theorem. This is fixable by reframing the claims, adding a limitations statement, and either providing a convergence or suboptimality analysis or explicitly labeling the method heuristic.\n\nThe citation pattern is heavy on the author's own prior work, but those papers are directly relevant and the reliance is not circular. No code or data is released, which limits reproducibility. This is a solid empirical paper that deserves refereeing, but only after the claims are brought in line with what is actually shown.","headline":"A practical hybrid that delivers real speedups and strong deblurring, but the abstract overclaims accuracy and theoretical guarantees that the paper's own convergence section disclaims.","tokens_in":20047,"tokens_out":2992,"would_cite":false,"duration_ms":28311,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68U10","65K10","68T07","90C26"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that alternating a pretrained neural network's guess with a few model-based iterations solves non-convex TpV medical imaging problems faster and more stably than either approach alone.","keywords":["non-convex optimization","total p-variation","incremental reweighted algorithm","deep learning initialization","medical image deblurring","sparse-view CT","ground-truth-free training","hybrid model-based reconstruction"],"falsifier":"Take the incDG test sets, replace the network output at every stage with either the raw current iterate or a fixed blurred image, and measure RE/SSIM: if the gap to incTpV vanishes, the claimed advantage is due to the neural initialization rather than the incremental mechanism. Alternatively, perturb the scheduler parameters $\\alpha_p$ and the iteration budget $K$ slightly and check whether final RE/SSIM jumps discontinuously, which would indicate the result is a tuned artifact of the schedule.","tokens_in":18971,"feed_emoji":"🩻","tokens_out":8752,"duration_ms":87113,"temperature":0.7,"pith_summary":"Medical image reconstruction from blurred or subsampled measurements is posed here as a non-convex inverse problem whose regularizer, the total $p$-variation ($\\mathrm{TpV}$) norm of the gradient image, approximates the $\\ell_0$ quasi-norm and therefore has many local minima. The paper claims that incDG, a hybrid algorithm alternating a pretrained network's refined guess with a few model-based iterations per stage, efficiently approximates the $\\ell_0$-optimal solution of this problem. The network provides a strong initialization, while the iterative solver keeps the reconstruction consistent with the measured data and stabilizes the output. In deblurring and sparse-view CT experiments, incDG reports lower relative error and higher SSIM than a purely iterative incremental solver and a purely network-based variant, and it keeps most of that quality when trained without ground truth. If true, this gives medical imaging a way to combine the speed of deep learning with the reliability of model-based optimization.","feed_headline":"AI-guided optimizer beats deep learning and classical solvers","feed_subtitle":"Non-convex TpV reconstruction stays stable and fast, even when trained without ground truth.","key_machinery":"The central object is the incremental homotopy schedule of subproblems, where the rule $p^{(h+1)} = p^{(h)}\\alpha_p$ pushes the prior from convex total variation ($p=1$) toward the $\\ell_0$ quasi-norm, and $\\lambda^{(h+1)} = \\lambda^{(h)} f^{(h)}/f^{(h-1)}$ adapts the regularization weight from the objective's successive values. Each stage is solved by an iterative reweighted $\\ell_1$ majorization of the non-convex $\\mathrm{TpV}$ term, with weights $w_i = p/(|Dx|_i^{1-p}+\\xi)$, using only five Chambolle-Pock iterations; a ResUNet (a U-shaped convolutional network with residual skip connections) maps the previous iterate to a refined guess before those iterations. The mechanism works because the network lowers the cost of each stage, the majorization keeps each subproblem convex, and the schedule gradually removes regularization while sharpening the sparsity prior.","core_discovery":"On its own terms, the paper establishes a constructive claim: the $\\ell_0$-optimal solution of a $\\mathrm{TpV}$-regularized imaging problem can be approximated by running a sequence of cheap convex subproblems, each initialized by a convolutional network and refined by five Chambolle-Pock iterations, while the sparsity parameter $p$ is driven geometrically from 1 toward 0 and the regularization weight $\\lambda$ is adapted from the ratio of successive objective values. The resulting incDG algorithm is reported to reach deblurring and CT reconstructions with better RE and SSIM than the purely model-based incTpV and the purely learned incNN, and with far fewer iterations than incTpV (20-30 versus 270-3100). The paper further claims that the model-based refinement is what confers stability: incNN's SSIM fluctuates widely and it hallucinates streaking artifacts, whereas incDG's metric spread stays narrow. Finally, training the networks against incTpV solutions instead of ground truth degrades quality only slightly, which the paper offers as evidence that the method can be deployed where clean references do not exist.","pith_inferences":["The same architecture could likely accelerate other non-convex sparsity priors (log-sum, atan, or $\\ell_p$ with different $p$ paths), since the machinery only needs a convex majorizer and a network that maps current iterates toward the target; the paper demonstrates only $\\mathrm{TpV}$.","Because most of incDG's runtime is fixed overhead (network forward pass and initialization), a lighter network or fused execution could push reconstruction toward real-time without changing the optimization logic.","The per-stage networks are trained independently with an MSE loss; a testable extension is to train all cascaded networks jointly with a loss on the final output, which could exploit the cascade structure and possibly improve end-to-end accuracy.","The stability claim implies a concrete diagnostic: under distribution shift or adversarial perturbation, incDG's worst-case SSIM should degrade gracefully while incNN's drops sharply; measuring that gap would validate whether the model-based iterations genuinely cap hallucination."],"forward_implications":["A model-based solver can run in a clinical time budget without losing accuracy: incDG finishes CT reconstruction in about 4.5 seconds versus 53 seconds for the incremental model-based solver, with comparable or better image quality.","A short model-based refinement protects any learned reconstruction against the instability that plagues standalone networks: incNN's SSIM swings widely and produces hallucinated streaks, while incDG's metrics stay concentrated.","Ground-truth-free training is a viable route: using model-based incremental solutions as training targets yields reconstructions almost as good as supervised training, so the method can be applied when no clean anatomical images are available.","The incremental schedule itself improves on fixed-parameter $\\mathrm{TpV}$ solutions, producing brighter, better-contrasted images that preserve low-contrast boundaries, which is the clinically relevant goal of enhancing subtle findings."],"supporting_citations":[{"why":"Supplies the incremental scheme with automatic $\\lambda$ and $p$ updating that incDG adapts to $\\mathrm{TpV}$.","marker":"[33]"},{"why":"Introduces the Deep Guess strategy of using a neural network's output as the starting point for a model-based solver.","marker":"[35]"},{"why":"Provides the primal-dual Chambolle-Pock algorithm used to solve each convex subproblem.","marker":"[16]"},{"why":"Establishes reweighted $\\ell_1$ minimization for sparsity enhancement, the foundation of the iterative reweighting step.","marker":"[13]"},{"why":"Gives the adaptive rule $\\lambda^{(h+1)} = \\lambda^{(h)} f^{(h)}/f^{(h-1)}$ for the regularization parameter.","marker":"[41]"},{"why":"Supplies the ground-truth-free training protocol where variational solutions serve as network targets.","marker":"[21]"},{"why":"Motivates the weighted $\\ell_1$ subproblem formulation for gradient-sparse CT reconstruction.","marker":"[50]"},{"why":"Supplies the convergence-of-descent background for semi-algebraic objectives used to justify the incremental penalization approach.","marker":"[4]"}],"fun_headline_variants":["Hybrid solver beats pure AI and classical methods","AI-guided refinement wins in non-convex imaging","Fast stable non-convex solver outperforms deep learning","No ground truth needed: AI+model solver excels","incDG: AI plus incremental optimization for sharper scans"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a handful of inexact inner iterations (five per stage), guided by the network's guess and the adaptive schedule, lands in a good local minimum of the non-convex $\\mathrm{TpV}$ objective; the paper explicitly notes that convergence of this inexact incremental algorithm is not proven.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid solver beats pure AI and classical methods","AI-guided refinement wins in non-convex imaging","Fast stable non-convex solver outperforms deep learning","No ground truth needed: AI+model solver excels","incDG: AI plus incremental optimization for sharper scans"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000177,"raw_usage":{"total_tokens":1322,"prompt_tokens":1006,"completion_tokens":316,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":622,"completion_tokens_details":{"reasoning_tokens":242}},"tokens_in":622,"tokens_out":316,"duration_ms":3668,"temperature":1.0,"reasoning_tokens":242,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:58:31.524164+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the incDG test sets, replace the network output at every stage with either the raw current iterate or a fixed blurred image, and measure RE/SSIM: if the gap to incTpV vanishes, the claimed advantage is due to the neural initialization rather than the incremental mechanism. Alternatively, perturb the scheduler parameters $\\alpha_p$ and the iteration budget $K$ slightly and check whether final RE/SSIM jumps discontinuously, which would indicate the result is a tuned artifact of the schedule.","supporting_citations":[{"cited_title":"A nonconvex penalization algorithm with automatic choice of the regularization parameter in sparse imaging","cited_arxiv_id":null,"evidence_quote":"Supplies the incremental scheme with automatic $\\lambda$ and $p$ updating that incDG adapts to $\\mathrm{TpV}$."},{"cited_title":"Deep guess acceleration for explainable image reconstruction in sparse-view ct","cited_arxiv_id":null,"evidence_quote":"Introduces the Deep Guess strategy of using a neural network's output as the starting point for a model-based solver."},{"cited_title":"A first-order primal-dual algorithm for convex problems with applications to imaging","cited_arxiv_id":null,"evidence_quote":"Provides the primal-dual Chambolle-Pock algorithm used to solve each convex subproblem."},{"cited_title":"Enhancing sparsity by reweighted L1 minimization","cited_arxiv_id":null,"evidence_quote":"Establishes reweighted $\\ell_1$ minimization for sparsity enhancement, the foundation of the iterative reweighting step."},{"cited_title":"An iterative l{1}-based image restoration algorithm with an adap- tive parameter estimation","cited_arxiv_id":null,"evidence_quote":"Gives the adaptive rule $\\lambda^{(h+1)} = \\lambda^{(h)} f^{(h)}/f^{(h-1)}$ for the regularization parameter."},{"cited_title":"RISING: A new framework for model- based few-view CT image reconstruction with deep learning","cited_arxiv_id":null,"evidence_quote":"Supplies the ground-truth-free training protocol where variational solutions serve as network targets."},{"cited_title":"Sidky and et al","cited_arxiv_id":null,"evidence_quote":"Motivates the weighted $\\ell_1$ subproblem formulation for gradient-sparse CT reconstruction."},{"cited_title":"Convergence of descent methods for semi-algebraic and tame problems: proximal algorithms, forward–backward splitting, and regularized gauss–seidel methods","cited_arxiv_id":null,"evidence_quote":"Supplies the convergence-of-descent background for semi-algebraic objectives used to justify the incremental penalization approach."}],"review_version":1}