{"id":"ea402c8e-4e9f-47ef-bf0b-452e7a10d8f7","arxiv_id":"2501.12245","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A ResUNet predicts per-pixel brightness and contrast parameters to automatically normalize radiographic X-ray images, achieving 24.75 dB PSNR and 0.8431 SSIM on a clinical test set.","lead":"This paper presents a deep learning model that adjusts X-ray image brightness and contrast using per-pixel settings that can be shown as explanation maps. A generalist might read it because inconsistent X-ray presentations across patients and scanners are a common clinical annoyance, and automating the correction could save time.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim rests on an unvalidated in-house tool as ground truth; without expert or clinical validation of the tool's parameters, both image quality and interpretable maps are unsupported.","rationale":"The reader's weakest_assumption identifies exactly the load-bearing concern: the in-house tool's parameter annotations are treated as ground truth without clinical validation. This is the most consequential vulnerability because it undermines both the quantitative evaluation and the interpretability claim. The proposed method is a plausible regression architecture, but its empirical conclusions depend entirely on the assumption that the tool's brightness and contrast settings match expert preferences. No code, data, or external validation is provided, so the assumption cannot be independently checked. The reader's conditional acceptance is appropriate: the paper describes an interesting engineering approach, but the central claim should not be accepted without a radiologist-based validation study of the ground truth. I see no other concern that is more load-bearing; the PSNR overclaim in the abstract is a secondary issue that could be corrected by rewording. Therefore, the verdict remains CONDITIONAL, and the reader's concern is affirmed.","tokens_in":3802,"tokens_out":2587,"duration_ms":27537,"concrete_test":"Select a held-out subset of 30 to 50 test images and have 3 to 5 radiologists independently perform their preferred global and regional brightness and contrast adjustments on a clinical workstation. Measure agreement between the in-house tool's parameters (a, b, wc, ww) and each radiologist's parameters using intraclass correlation, and compare tool-generated images with radiologist-adjusted images using PSNR/SSIM and a forced-choice preference test. If inter-radiologist agreement is low or the tool's parameters deviate systematically from expert adjustments, the ground truth is biased and the central claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's core assertion is that the proposed network produces clinically desirable brightness and contrast correction with interpretable parameter maps. For this to hold, the ground truth must represent expert-preferred presentation. In Section 3, all 429 images are annotated by an in-house engineering tool, and Eq. 1 supervises the network against both the tool-generated image (imgGT) and the tool's parameters (agt, bgt, wwgt, wcgt). The tool itself is never validated against radiologist adjustments or any independent clinical standard. If the tool's notion of 'consistent presentation' differs from what radiologists prefer, the network will learn a biased target, and the reported PSNR/SSIM values (24.75 dB, 0.8431) merely measure fidelity to the tool, not clinical quality. Moreover, the interpretable maps are not an independent explanation of enhancement: they are regressed directly to the tool's internal parameters, so any bias in the tool propagates into the claimed interpretability. The paper also overstates superiority: Table 1 shows ResUNet achieves higher PSNR (25.10 vs 24.75), with the proposal favored only via SSIM, but the central unresolved issue remains the unvalidated ground truth.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an interpretable deep-learning method for brightness and contrast correction of radiographic X-ray images. A ResUNet predicts per-pixel parameter maps (slope a, bias b, window center wc, window width ww), which are then used to remap the input image through regional and global LUT-style operations. The network is trained with an MSE loss that simultaneously supervises the reconstructed image and the four parameter maps against outputs of an in-house engineering tool. On a test set of 43 clinical X-ray images, the method reports PSNR 24.75 dB and SSIM 0.8431 at a downscale factor of 8, and the paper claims consistent presentation across anatomies and scanning positions with interpretable maps.","tokens_in":4121,"tokens_out":2572,"duration_ms":27520,"significance":"If the central claim holds, the method would be practically valuable: it automates a time-consuming manual presentation-correction step, works at full resolution with modest processing time, and provides parameter maps that are more interpretable than a black-box image-to-image network. The paper includes comparisons with several relevant baselines (ResUNet, DLBC, 3DLUT, Sup-DCE) and reports processing times. A clear strength is the explicit mapping from predicted parameters to a conventional brightness/contrast operation, which gives the method a transparent mechanism. However, the significance is conditional on the validity of the in-house tool used as ground truth; without expert or clinical validation, the reported PSNR/SSIM values only demonstrate fidelity to that tool, and the interpretability claim inherits the tool's biases.","major_comments":[{"comment":"The central training and evaluation target is the output of an in-house engineering tool, with no validation of that tool against radiologist adjustments or any independent clinical standard. Eq. (1) supervises both the reconstructed image and the four parameter maps directly against the tool's outputs, so the network can only learn the tool's notion of 'consistent presentation.' The reported PSNR and SSIM therefore quantify fidelity to the tool, not clinical quality, and the interpretable maps are not independent explanations but regressed reproductions of the tool's internal parameters. The authors should provide expert validation, a reader study, or at least a detailed characterization of the tool's agreement with clinical preferences.","section":"Section 2, Eq. (1) and Section 3"},{"comment":"The statement that 'The proposal has the best performance' is contradicted by the same table, in which ResUNet achieves a higher PSNR (25.10 dB vs. 24.75 dB). The paper explains that ResUNet has blurrier artifacts and lower SSIM, which may be true, but the claim of overall superiority needs a quantitative basis, such as a perceptual metric, a significance test, or a clearly stated preference for SSIM over PSNR. As written, the comparison does not support the global 'best performance' assertion.","section":"Section 3, Table 1"},{"comment":"The claim of consistent brightness and contrast presentation across different images is supported only by qualitative inspection of Figure 4. No metric is defined for presentation consistency, no statistical comparison is made between the proposed method and the baselines, and no error bars or confidence intervals are reported. The authors should quantify consistency, for example by measuring the variance of presentation-relevant statistics (e.g., window center, window width, mean intensity in anatomical regions) across test images and across methods, and should report the spread of PSNR and SSIM results over the 43 test images.","section":"Section 3, Figure 4 and consistency claim"},{"comment":"The loss function in Eq. (1) sums five MSE terms with equal weighting, but the relative scales of image intensities and parameter values are very different, and no ablation or justification for this weighting is provided. Since the parameter-map losses directly shape the interpretable maps, a poorly chosen weighting could dominate training and distort the claimed interpretability. The authors should report the loss weights, analyze their sensitivity, or provide evidence that the chosen weighting does not materially affect the results.","section":"Section 2, Eq. (1) and Section 3, dataset annotation"}],"minor_comments":[{"comment":"There is a typo in 'donwsample factor' (should be 'downsample factor'), and 'structure similarity index measurement' should be 'structural similarity index measure.'","section":"Section 3"},{"comment":"The phrase 'accuracy of 24.75 dB PSNR' is imprecise; PSNR is a distortion or quality metric, not an accuracy measure. Please rephrase, e.g., 'achieving PSNR of 24.75 dB and SSIM of 0.8431.'","section":"Abstract and Section 4"},{"comment":"The remapping formula that converts predicted a, b, wc, and ww into the enhanced image is described verbally as a piecewise linear function and a sigmoid-like function, but the exact equations are not given. Providing the mathematical formulation would improve reproducibility and make the interpretability argument more concrete.","section":"Section 2, Figure 2"},{"comment":"The downscale factor is described as a memory-saving choice, but the paper does not discuss how the upsampling of parameter maps affects spatial resolution of local adjustments; a sentence on this trade-off, or an ablation on a smaller set, would help.","section":"Section 3"},{"comment":"The in-house engineering tool is not described beyond 'efficiently annotate the images with correct parameters.' Some algorithmic description or reference is needed for readers to assess what kind of 'correct' presentation the tool encodes and to allow independent reproduction.","section":"Section 3"}],"recommendation":"major_revision","confidential_remarks":"The paper's ground-truth generation relies entirely on an in-house GE Healthcare tool, and the annotation tool itself is not publicly described or validated. This is a reproducibility and objectivity concern: the authors should be asked to provide a detailed algorithm description, validation against expert-adjusted images, or a release path. The fit with a general medical imaging audience is appropriate, but the central claim hinges on external validation of that tool."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: the per-pixel slope/bias LUT plus global window center/width map prediction is a real, if incremental, architecture extension, and the paper is honest enough to show the comparison table. But the load-bearing assumption is that an in-house engineering tool defines 'correct presentation,' and that tool is never validated against radiologist expertise. Without that, the PSNR/SSIM numbers measure fidelity to the tool, not clinical quality.\n\nWhat's new: prior work predicts global window/level (DLBC), image-level LUTs (3DLUT), or global curves (Zero-DCE). Here each pixel gets its own slope and bias plus global window parameters, supervised directly on the parameter maps (Eq. 1). That is a sensible way to make the enhancement locally adaptive while keeping the mapping interpretable by construction. The downscale study (4/8/16) is a nice practical touch showing robustness to low-res parameter maps. The qualitative regional detail improvements over global-only methods are credible.\n\nThe main soft spot is the ground truth. All 429 images are annotated with an in-house tool, and Eq. 1 trains against both the tool's image output and its internal parameters. The tool is described as 'efficient' but not validated against radiologists or any independent standard. So the interpretability claim is weaker than it looks: the maps are regressed to the tool's parameters, not to expert-verified corrections; they explain what the network did relative to the tool, not why the enhancement is clinically good. Second, Table 1 contradicts the abstract's 'best performance' claim: ResUNet gets higher PSNR, and the proposed method wins only on SSIM. No error bars, no significance tests, and a 43-image test set. That is a fixable but real problem. Third, the consistency claim in Figure 4 is qualitative; no metric quantifies it. No code or data is released, which limits reproducibility. None of these are fatal; the network does what the loss asks.\n\nWho this is for: researchers working on automatic X-ray presentation correction or interpretable image enhancement. They will find a clear baseline and a useful discussion of parameter-map supervision. The central argument holds up as an engineering contribution; the clinical usefulness remains unproven.\n\nRecommendation: send it to peer review. The method is sound and the flaw is fixable, but the authors should be asked to validate the tool against expert-adjusted images (or at least report a reader study) and to moderate the superiority claim.","headline":"A plausible engineering contribution whose central 'quality' claim is anchored to an unvalidated in-house tool; needs clinical validation before the consistency story holds.","tokens_in":4605,"tokens_out":2183,"would_cite":false,"duration_ms":22295,"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":"A deep network enhances radiographic X-ray images by predicting per-pixel brightness and contrast parameters, producing interpretable maps that explain the enhancement.","keywords":["radiographic images","image quality enhancement","image consistency","deep learning","interpretable mapping","brightness and contrast correction","pixel-wise lookup table","ResUNet"],"falsifier":"A clinical reader study in which radiologists view triples of the original image, the in-house tool's annotated output, and the network's enhanced output across diverse anatomies and patient sizes, and state which presentation they prefer, would settle whether the tool's definition of consistent presentation is the correct target; if experts do not systematically prefer the tool's annotation, the ground truth the network learns is falsified.","tokens_in":3655,"feed_emoji":"🩻","tokens_out":8402,"duration_ms":72911,"temperature":0.7,"pith_summary":"Radiographic X-ray images often reach radiologists with inconsistent brightness and contrast because of patient size, position, and scanning protocol, forcing manual adjustment. This paper proposes a deep network that predicts per-pixel brightness and contrast parameters—slope, bias, window center, and window width—and uses those parameters to remap the input image into a consistently presented output. Because the remapping follows the conventional brightness-and-contrast workflow, each pixel's adjustment is an interpretable map rather than a black-box change. On a clinical dataset of 429 images the method reaches 24.75 dB PSNR and 0.8431 SSIM at full resolution, and the authors argue the interpretable maps let clinical experts see why each region was enhanced.","feed_headline":"Enhancement network hits 24.75 dB PSNR on X-rays","feed_subtitle":"Per-pixel brightness and contrast maps show clinicians why each region is adjusted.","key_machinery":"The central object is a pixel-wise lookup table parametrized by four maps: a slope $a$ and bias $b$ for the regional piecewise-linear adjustment, and a window center $wc$ and window width $ww$ for the global sigmoid-like adjustment. A five-level ResUNet with residual convolution blocks predicts these maps from an input downsampled by factor 8; the maps are upsampled to full resolution and applied through a pixel-level remapping formula to produce the enhanced image. The training loss is the sum of image-level mean squared error and parameter-level mean squared error for each of the four maps, so the network is explicitly supervised to reproduce the annotation tool's parameter choices, not just the final pixel values.","core_discovery":"The central claim is that a ResUNet trained to output per-pixel lookup-table parameters can perform both global and regional brightness/contrast correction in one forward pass, and that the parameter maps themselves are the explanation of the enhancement. The objective loss combines image-level mean squared error with direct mean squared error on each parameter map, where the ground-truth parameters come from an in-house engineering tool that defines consistent presentation. The paper reports the proposed method reaches 24.75 dB PSNR and 0.8431 SSIM at downscale factor 8 on full-resolution test images, and that it outperforms the DLBC, 3DLUT, and Sup-DCE baselines overall. A plain ResUNet has slightly higher PSNR but noticeably lower SSIM and visibly blurrier output, which the authors attribute to its upsampling design.","pith_inferences":["A clinical preference study is the natural next test: the entire method inherits the in-house tool's definition of consistent presentation, and if expert taste diverges from that definition, the interpretable maps would reveal systematic disagreement per anatomy.","Because window center and width are standard concepts in CT and other modalities, the same four-parameter remapping could be transferred to CT presentation consistency, where interpretability would be even more natural.","If the annotation source is replaced by a self-supervised or unsupervised consistency objective, the architecture itself could be reused unchanged; this would test whether the parametrization or the supervised target is the main driver of the reported quality.","The decisive comparison may be clinical rather than numerical: plain ResUNet has higher PSNR but much lower SSIM, so the practical case for this method rests on whether the interpretable maps change how clinicians review and trust the output."],"forward_implications":["Radiologists would no longer need to manually adjust window/level or regional equalization for each study, since the network supplies consistent global and regional correction automatically.","Because the adjustment is expressed through interpretable parameter maps, a clinician can audit exactly which regions were brightened or darkened and by how much, which is impossible with end-to-end image translation.","The low sensitivity to downscale factor (24.36, 24.75, and 24.89 dB PSNR for factors 16, 8, and 4) indicates that coarse parameter maps suffice for full-resolution enhancement, easing memory and deployment constraints.","The interpretable maps separate presentation correction from pixel generation, so downstream quality assurance could verify the enhancement by inspecting the maps rather than re-running the network."],"supporting_citations":[{"why":"Defines the deep-learning brightness and contrast correction baseline that the paper re-implements and compares against.","marker":"[1]"},{"why":"Provides the conventional equalized-contrast display processing chain that motivates the regional and global adjustment formulation.","marker":"[2]"},{"why":"Supplies the general framework for optimizing display contrast, detail, and noise that underlies the interpretable remapping design.","marker":"[3]"},{"why":"Supplies the U-Net encoder-decoder architecture used as the backbone of the proposed parameter map predictor.","marker":"[4]"},{"why":"Supplies the residual identity-mapping blocks used in the five-level ResUNet to train the parameter prediction.","marker":"[5]"},{"why":"Provides the image-adaptive 3D lookup table enhancement baseline used for performance comparison.","marker":"[6]"},{"why":"Provides the supervised Zero-DCE baseline used as a low-light curve-estimation comparison.","marker":"[7]"}],"fun_headline_variants":["Interpretable mapping hits 24.75 dB PSNR on X-ray enhancement","Explainable per-pixel maps correct X-ray brightness and contrast","Global and local X-ray brightness fixes now have clear explanations","Model outputs per-pixel maps that explain X-ray enhancement choices","24.75 dB PSNR and interpretable maps for consistent X-ray reads"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the in-house engineering tool's parameter annotations define the correct consistent presentation; if that tool's notion of brightness and contrast does not match radiologist preference, both the enhanced images and the interpretable maps inherit the bias.","fun_headline_variants_meta":{"raw":{"variants":["Interpretable mapping hits 24.75 dB PSNR on X-ray enhancement","Explainable per-pixel maps correct X-ray brightness and contrast","Global and local X-ray brightness fixes now have clear explanations","Model outputs per-pixel maps that explain X-ray enhancement choices","24.75 dB PSNR and interpretable maps for consistent X-ray reads"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000556,"raw_usage":{"total_tokens":2629,"prompt_tokens":912,"completion_tokens":1717,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":1624}},"tokens_in":528,"tokens_out":1717,"duration_ms":13741,"temperature":1.0,"reasoning_tokens":1624,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:21:32.382711+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A clinical reader study in which radiologists view triples of the original image, the in-house tool's annotated output, and the network's enhanced output across diverse anatomies and patient sizes, and state which presentation they prefer, would settle whether the tool's definition of consistent presentation is the correct target; if experts do not systematically prefer the tool's annotation, the ground truth the network learns is falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the deep-learning brightness and contrast correction baseline that the paper re-implements and compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the conventional equalized-contrast display processing chain that motivates the regional and global adjustment formulation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the general framework for optimizing display contrast, detail, and noise that underlies the interpretable remapping design."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the U-Net encoder-decoder architecture used as the backbone of the proposed parameter map predictor."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the residual identity-mapping blocks used in the five-level ResUNet to train the parameter prediction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the image-adaptive 3D lookup table enhancement baseline used for performance comparison."},{"cited_title":"C., ``Learning to enhance low-light image via zero-reference deep curve estimation,'' IEEE transactions on pattern analysis and machine intelligence 44 (8), 4225--4238 (2021)","cited_arxiv_id":null,"evidence_quote":"Provides the supervised Zero-DCE baseline used as a low-light curve-estimation comparison."}],"review_version":1}