{"id":"d8853c68-6bb4-40be-a811-02a00e2e2a18","arxiv_id":"1908.06965","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"Fixed-Point GAN adds a conditional identity loss to StarGAN, preserving same-domain images and enabling disease detection and localization by subtracting a generated healthy image from the input.","lead":"Adding a 'leave unchanged when asked for no change' rule to a popular image-translation GAN cuts unintended edits and lets the network remove diseases or objects from images. By subtracting the original image from the 'healed' one, the method finds and locates disease using only image-level labels, which are cheap to collect.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim rests on an unenforced minimal-transformation premise; the paper's own real-BRATS low-FP results contradict the unqualified superiority claim.","rationale":"The reader's weakest assumption identifies the load-bearing issue: the disease detection and localization framework works only if the generator performs a minimal transformation that removes disease while preserving all healthy content. I agree, and would sharpen it: this premise is not merely unproven; it is only indirectly encouraged by the loss design and is contradicted by the paper's own real-BRATS low-false-positive results. The conditional identity loss applies only when the target domain equals the source domain, so it does not directly constrain what changes during cross-domain translation. Cycle consistency can be satisfied by non-minimal transformations that are undone by the reverse translation. Thus the difference map has no formal guarantee of sparsity or disease specificity. The residual/delta map helps reduce the magnitude of changes but does not restrict them to lesion pixels. The paper's caveats about partial removal of large lesions and underperformance at low false-positive rates are symptoms of this same missing premise. That said, the paper has genuine strengths: the implementation is public, the ablation in Table 4 shows that fixed-point translation learning accounts for most of the improvement over StarGAN, and the proposed training scheme is a reasonable extension of StarGAN. The conditional verdict remains appropriate; the headline claims should be narrowed to the operating regimes actually supported by the evidence.","tokens_in":15908,"tokens_out":5900,"duration_ms":64397,"concrete_test":"On the BRATS 2013 synthetic test set (which has pixel-level lesion masks), run the released Fixed-Point GAN code and compute, for each diseased slice, the mean absolute difference between the input and the generated healthy image inside the ground-truth lesion mask versus inside the healthy-tissue mask. If the outside-lesion change is not a small fraction (e.g., below 25%) of the inside-lesion change, the minimal-transformation premise is refuted and the difference map cannot be trusted as a disease-specific localizer. This directly tests whether the signal used for detection and localization is disease-specific rather than an artifact of non-minimal normal-structure changes.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on the premise, entered in Sec. 4.2, that the generator performs a minimal, disease-only transformation, so subtracting the generated healthy image from the input reveals the lesion. This premise is not enforced by the losses. The conditional identity loss (Eq. 5) supervises only same-domain translations, and the cycle-consistency loss (Eq. 4) penalizes round-trip error, not per-pixel changes outside the target attribute. A non-minimal cross-domain map that also alters healthy structures can satisfy both losses if the reverse translation restores the original image. The residual/delta architecture (tanh(G(x,c)+x), Appendix Implementation Details) biases toward small deltas, but small does not mean confined to lesion pixels. The paper's own results show the premise fails in important regimes: Sec. 4.4 concedes that large objects such as sunglasses and brain lesions are only partially removed, and Sec. 4.2 states that on real BRATS images the method is outperformed at the low false-positive range. Since image-level detection takes the maximum of the difference map and localization applies connected-component analysis to the same map, any residual normal-structure change is scored as disease evidence. The abstract's unqualified 'outperforms' and 'surpasses' claims are therefore supported only if the minimal-transformation premise holds quantitatively, which the paper does not demonstrate.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Fixed-Point GAN, an extension of StarGAN that trains a single generator for both cross-domain and same-domain translations. The fixed-point training scheme adds a conditional identity loss for same-domain translation, revises the adversarial and domain-classification losses to cover both same- and cross-domain outputs, and applies cycle consistency to both. The generator is implemented with a residual/delta architecture. The authors evaluate multi-domain translation on CelebA and use the difference between an input image and its translation to the healthy domain for weakly supervised disease detection and localization on BRATS 2013 brain MRI and a pulmonary embolism CTPA dataset, comparing against StarGAN, CAM, ResNet-50-CAM, f-AnoGAN, and Alex et al. The central claim is that the proposed training yields minimal, fixed-point translations, leading to state-of-the-art image translation and superior disease detection/localization with only image-level labels.","tokens_in":16166,"tokens_out":5321,"duration_ms":48635,"significance":"The idea is appealing: a GAN that can 'virtually heal' an image and reveal disease by subtraction would make weakly supervised localization practical, since it requires only image-level labels. The paper includes a useful ablation (Tab. 4) separating the delta-map and fixed-point-training contributions, and the fixed-point component clearly drives most of the gain. The authors also make the implementation publicly available. However, the strongest claims ('outperforms the state of the art', 'surpasses predominant weakly-supervised localization methods') are not fully supported by the evidence as presented, because the minimal-transformation assumption is not quantitatively verified, comparisons are made at non-commensurate operating points, and no uncertainty estimates are given. With revisions that tighten these claims and add targeted analyses, the contribution would be a valuable advance.","major_comments":[{"comment":"The load-bearing minimal-transformation premise is stated but not enforced or quantitatively verified. Eq. (5) penalizes L1 only for same-domain translation, and Eq. (4) cycle consistency penalizes round-trip error; neither prevents the generator from modifying normal structures as long as the reverse map restores them. The paper's own observations—'neither StarGAN nor Fixed-Point GAN can completely remove large objects, like sunglasses or brain lesions' (Sec. 4.4) and 'our method is outperformed at the low false positive range' on real BRATS (Sec. 4.2)—show that the difference map is not guaranteed to be confined to disease. Since image-level detection uses the maximum of the difference map and localization uses connected-component analysis, any residual normal-structure change contributes to false positives. The abstract's unqualified superiority claim should be replaced with a claim conditional on the minimality of the learned transformation, and the paper should report a direct measure of preservation, e.g., the magnitude of the difference map in regions known to be healthy or the per-pixel specificity outside the lesion ground truth.","section":"Sec. 4.2, Sec. 4.4, Eq. (4), Eq. (5)"},{"comment":"The empirical support lacks uncertainty quantification and significance testing. AUC, FROC sensitivity, and IoU are reported as point estimates (e.g., mean IoU 0.2609±0.1283 vs 0.3483±0.2420, AUC 0.9668 vs 0.8832), with no confidence intervals or tests, and the test sets are small (6 real BRATS patients; 127 unique PEs). Many reported gaps, especially those with overlapping standard deviations, could easily arise from noise. Report bootstrap confidence intervals or significance tests, and show error bars on the ROC/FROC curves.","section":"Sec. 4.2, Tab. 4, Fig. 4"},{"comment":"The evaluation protocol for f-AnoGAN is selected after inspecting test performance: 'we use the average activation of difference maps as the detection score, because we find it more effective than using the maximum activation ... and also more effective than the anomaly scores proposed in the original work.' Choosing the scoring function based on the test set invalidates the comparison as a fair head-to-head. State the scoring rule before evaluation, or report all considered scores with a multiple-comparison correction, and apply the same selection procedure to all methods.","section":"Sec. 4.2, f-AnoGAN comparison"},{"comment":"The head-to-head claim against ResNet-50-CAM is made at different operating points: Fixed-Point GAN achieves 84.5% sensitivity at 1 false positive per image, while ResNet-50-CAM achieves 60% at 0.037 false positives per image. These numbers are not comparable. Report sensitivity for all methods at a common false-positive rate, or use a partial-area FROC measure over the same FP range; the same issue appears in the real-image FROC (Fig. 4d) and the PE FROC (Fig. 5b).","section":"Sec. 4.2, Fig. 4b, Fig. 4d, Fig. 5b"},{"comment":"The same-domain L1 comparison is partly by construction. Eq. (5) directly minimizes ||G(x,c_x)-x||_1, and Tab. 3 evaluates exactly this quantity. The fact that Fixed-Point GAN has lower same-domain L1 than StarGAN is therefore a sanity check of the loss rather than independent evidence of translation quality. The paper should present same-domain L1 as a training objective check, and use the independently trained classifier (Tab. 2) and visual inspection as the primary translation-quality evidence; ideally also report a perceptual or human-evaluation metric.","section":"Sec. 4.1, Eq. (5), Tab. 3"}],"minor_comments":[{"comment":"The abstract should be tempered to acknowledge the low-false-positive-range underperformance on real BRATS and the partial-removal failures on large objects, both of which are disclosed in the body of the paper.","section":"Abstract, Sec. 4.2, Sec. 4.4"},{"comment":"The main text defines the adversarial loss with the standard log form in Eq. (1), while the appendix states that the actual objective is the Wasserstein loss with gradient penalty in Eq. (8). Clarify which objective is used and keep the notation consistent throughout.","section":"Sec. 3, Appendix Implementation Details"},{"comment":"For the PE dataset, 'localization' is defined as correct image-level classification of the candidate patch, not as lesion-level localization via connected components; the text should describe this as candidate-level detection rather than localization to avoid overstating the result.","section":"Sec. 4.3"},{"comment":"Classification accuracy on generated images is an indirect quality measure and can be biased because the classifier was trained on real images; consider reporting an additional metric, such as a classifier trained on generated images or a perceptual similarity score.","section":"Tab. 2"},{"comment":"There is a duplicated word in 'sensitivity levels of of 88.9%' in Sec. 4.3, and the code URL is given as https://github.com/jlianglab/Fixed-Point-GAN in the abstract but as http://github.com/jlianglab/Fixed-Point-GAN in the implementation details; these should be unified.","section":"Sec. 4.3, Implementation Details"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is likely publishable after revision if the authors add uncertainty quantification, redo the f-AnoGAN and ResNet-CAM comparisons at common operating points, and temper the abstract. The clinical claim is potentially important, but the current presentation overstates certainty."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's real contribution is a clean, small modification of StarGAN: a conditional identity loss that explicitly trains same-domain translation, plus revised adversarial/domain/cycle losses. That fixes a real artifact problem in StarGAN, and the ablation in Tab. 4 makes a convincing case that the gain is mostly from that loss, not from the delta-map architecture. The detection-by-removal framework is also a nice transfer of the same idea to weakly supervised localization, and the fact that it works with image-level labels only is practically valuable. The paper is clearly written, the math is straightforward, and the qualitative figures support the basic mechanism.\n\nWhere I part ways with the authors is the overclaiming. The abstract says the method 'outperforms the state of the art' and 'surpasses predominant weakly-supervised localization methods' without qualification. The evidence does not support that unqualified statement. On real BRATS, the paper itself concedes it is 'outperformed at the low false positive range.' The localization numbers are FROC points at 1 false positive per image, not full curves, and the f-AnoGAN baseline's detection score was chosen post hoc after finding it worked better. There are no confidence intervals or significance tests anywhere. The reader's stress-test note about the minimal-transformation premise is fair: the losses encourage but do not enforce that the generator only changes disease pixels, and the paper's own Sec. 4.4 admission that large objects are only partially removed means the difference map will sometimes miss parts of the lesion and occasionally fire on normal structure.\n\nI don't think the central idea collapses. The method works well enough for detection and coarse localization on these datasets, and the PE results are strong. The claim just needs to be scaled back to 'competitive with or better than the compared baselines on the tested datasets under these operating points.' That is a publishable claim.\n\nThe citation pattern looks honest. They compare against the right baselines, use public implementations where available, and release code and data splits. No red flags.\n\nWho is this for? Anyone working on weakly supervised medical image localization or on reducing artifacts in multi-domain translation. It deserves a serious referee—the empirical gaps are fixable with proper statistical reporting and a more careful framing. I would send it out, but I would tell the authors to tone down the abstract and add error bars.","headline":"A genuine, modestly-scoped extension of StarGAN with an honest but self-admittedly incomplete evaluation; the detection-by-removal idea is clever and worth engaging, but the 'state of the art' claim outruns the evidence.","tokens_in":16727,"tokens_out":603,"would_cite":true,"duration_ms":8411,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Fixed-Point GAN trains a generator to translate any medical image into a healthy one, so subtracting the two reveals diseased regions.","keywords":["fixed-point translation","image-to-image translation","generative adversarial networks","weakly-supervised localization","disease detection","virtual healing","medical image analysis","difference map"],"falsifier":"Inject a synthetic lesion of known location and size into a healthy MRI, translate the image to the healthy domain, and inspect the difference map; if noticeable differences appear outside the injected lesion, or if healthy anatomy is modified, the minimal-transformation assumption is false.","tokens_in":15707,"feed_emoji":"🩺","tokens_out":8135,"duration_ms":71940,"temperature":0.7,"pith_summary":"Fixed-Point GAN asks whether a generative network can remove an object or disease from an image while preserving everything else. The paper's central claim is that supervising same-domain translation with a conditional identity loss, paired with revised adversarial, domain classification, and cycle consistency losses for cross-domain translation, produces a generator that changes only the minimal set of pixels required for translation. If that holds, any medical image, diseased or healthy, can be mapped to a healthy image, and subtracting the two exposes the disease. The paper reports that this fixed-point translation improves multi-domain image-to-image translation on faces and outperforms weakly-supervised localization and anomaly-detection baselines for brain lesions and pulmonary embolism using image-level labels only.","feed_headline":"A GAN that 'heals' medical images to reveal disease","feed_subtitle":"Translating any scan into a healthy version yields a difference map that localizes lesions from image-level labels alone.","key_machinery":"The load-bearing object is fixed-point translation learning. It adds a conditional identity loss $\\mathcal{L}_{id} = \\mathbb{E}_{x, c_x}[\\|G(x, c_x)-x\\|_1]$ that penalizes any change when the target domain is the source domain, and it revises the adversarial, domain classification, and cycle consistency losses so that same-domain translations are included; the cycle loss $\\mathcal{L}_{cyc} = \\mathbb{E}_{x, c_x}[\\|G(G(x, c_x), c_x)-x\\|_1]$ forces same-domain round trips to return the input. The generator also outputs a residual delta map that is added to the input before a $\\tanh$ activation, which makes it structurally easier to leave the image unchanged. Together these components enforce that only disease- or attribute-related pixels are modified.","core_discovery":"The discovery is a training scheme that endows a GAN with fixed-point translation: when the target domain equals the source domain, the generator is trained to behave as identity, $G(x, c_x) \\approx x$, and when the domains differ, it is trained through losses that also cover same-domain pairs, so it learns minimal transformations that avoid touching unrelated content. In the 'virtual healing' application, the generator always translates toward the healthy domain, and the absolute difference between the input and the generated healthy image—the difference map—is used to detect and localize disease. This yields a weakly supervised detection and localization pipeline that needs only image-level healthy/diseased labels.","pith_inferences":["The authors note that large objects such as sunglasses and brain lesions are only partially removed, so the difference map is a localization signal rather than a segmentation mask; complete removal would be needed for pixel-level segmentation, which remains open.","A natural extension is to feed the difference maps back as pseudo-labels to train a supervised segmentation network, potentially improving boundary accuracy.","Since the conditional identity loss applies only when target equals source, the trained generator is source-domain-independent by construction; this suggests the 'translate to healthy' operation could be applied to any input modality without changing the inference procedure.","Distribution shift is a testable risk: diseases or artifacts not seen during training could make the generator either fail to remove them or hallucinate changes; a healthy-only calibration set would reveal this failure mode."],"forward_implications":["Healthy images pass through the generator nearly unchanged, so their difference maps are clean and rarely trigger false positives.","Cross-domain translations, such as changing hair color or removing a lesion, are regularized to modify only task-relevant pixels, which improves translation quality.","Disease detection and localization can be carried out with only image-level labels by applying color quantization and connected-component analysis to the difference map.","The same fixed-point training transfers from natural images to brain MRI and CT pulmonary angiography without architectural changes.","Because the target domain is specified at inference and the source domain is not needed, the method works for images with unknown health status."],"supporting_citations":[{"why":"Baseline multi-domain image-to-image translation method that Fixed-Point GAN extends and compares against.","marker":"[8]"},{"why":"Sourced the cycle consistency loss idea that Fixed-Point GAN revises for same-domain translations.","marker":"[39]"},{"why":"Class activation map baseline for weakly-supervised localization that Fixed-Point GAN is compared against.","marker":"[37]"},{"why":"f-AnoGAN anomaly detection baseline used for brain lesion and pulmonary embolism comparisons.","marker":"[24]"},{"why":"GAN-based brain lesion detection baseline used for comparison.","marker":"[1]"},{"why":"ResNet-50 classifier used to compute CAM baselines for localization.","marker":"[11]"},{"why":"CelebA dataset used for multi-domain image-to-image translation evaluation.","marker":"[20]"},{"why":"BRATS 2013 brain tumor dataset used for lesion detection and localization experiments.","marker":"[21]"},{"why":"Provides the Wasserstein objective on which the revised adversarial loss is based.","marker":"[2]"},{"why":"Supplies the gradient penalty used to stabilize the revised adversarial loss.","marker":"[10]"}],"fun_headline_variants":["GAN learns fixed-point translation to virtually heal scans","Fixed-Point GAN: healing images to spot disease","Training GANs to heal images for disease localization","Virtual healing: GAN reveals disease by translating scans","From healing images to disease detection: a new GAN"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline assumes the generator changes only disease-related pixels and preserves all other image content, so the difference between a scan and its healthy translation marks exactly the disease.","fun_headline_variants_meta":{"raw":{"variants":["GAN learns fixed-point translation to virtually heal scans","Fixed-Point GAN: healing images to spot disease","Training GANs to heal images for disease localization","Virtual healing: GAN reveals disease by translating scans","From healing images to disease detection: a new GAN"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000485,"raw_usage":{"total_tokens":2386,"prompt_tokens":934,"completion_tokens":1452,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":1385}},"tokens_in":550,"tokens_out":1452,"duration_ms":10413,"temperature":1.0,"reasoning_tokens":1385,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:54:23.004607+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inject a synthetic lesion of known location and size into a healthy MRI, translate the image to the healthy domain, and inspect the difference map; if noticeable differences appear outside the injected lesion, or if healthy anatomy is modified, the minimal-transformation assumption is false.","supporting_citations":[{"cited_title":"Stargan: Uni- fied generative adversarial networks for multi-domain image-to-image translation","cited_arxiv_id":null,"evidence_quote":"Baseline multi-domain image-to-image translation method that Fixed-Point GAN extends and compares against."},{"cited_title":"Unpaired image-to-image translation using cycle- consistent adversarial networks","cited_arxiv_id":null,"evidence_quote":"Sourced the cycle consistency loss idea that Fixed-Point GAN revises for same-domain translations."},{"cited_title":"Learning deep features for discrimi- native localization","cited_arxiv_id":null,"evidence_quote":"Class activation map baseline for weakly-supervised localization that Fixed-Point GAN is compared against."},{"cited_title":"f-anogan: Fast unsupervised anomaly detection with generative adversarial networks","cited_arxiv_id":null,"evidence_quote":"f-AnoGAN anomaly detection baseline used for brain lesion and pulmonary embolism comparisons."},{"cited_title":"Generative adver- sarial networks for brain lesion detection","cited_arxiv_id":null,"evidence_quote":"GAN-based brain lesion detection baseline used for comparison."},{"cited_title":"Deep learning face attributes in the wild","cited_arxiv_id":null,"evidence_quote":"CelebA dataset used for multi-domain image-to-image translation evaluation."},{"cited_title":"The multimodal brain tumor image segmentation benchmark (brats)","cited_arxiv_id":null,"evidence_quote":"BRATS 2013 brain tumor dataset used for lesion detection and localization experiments."},{"cited_title":"Improved training of wasserstein gans","cited_arxiv_id":null,"evidence_quote":"Supplies the gradient penalty used to stabilize the revised adversarial loss."}],"review_version":1}