{"id":"910e76da-02c6-4d64-b6c7-b1f0d52cd1b8","arxiv_id":"1908.11324","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An anchor-free 3D CNN, formalizing lesions as center keypoints, outperforms its own anchor-based RPN on DeepLesion in FROC score (0.735 vs 0.708) and inference time (1.74s vs 1.95s/scan).","lead":"The authors propose a 3D anchor-free lesion detector that treats each lesion as a single center keypoint in CT scans, instead of using predefined anchor boxes. They report higher accuracy and faster inference than their own anchor-based baseline on the DeepLesion dataset, which is the largest public CT lesion dataset.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed anchor-free gain may rest on an under-tuned anchor-based baseline: the 9 hand-picked anchors in Section IV-B are never ablated or compared against any alternative configuration.","rationale":"The reader's weakest-assumption analysis identifies the same load-bearing point: the anchor-based baseline is configured once by hand with no tuning or sensitivity study, so the empirical comparison may be unfair. I agree that this is the most direct threat to the paper's central claim, because the two networks are otherwise matched, and anchor configuration is known to have a large effect on detection accuracy. The concern is concrete: the chosen anchor scales stop at 40 mm while lesions above 30 mm show a 8-point sensitivity gap between the two models, which is suspicious but not conclusive. The proposed test, k-means-derived anchors with the same anchor budget, directly addresses whether the baseline is weak. I also note that the evaluation criterion (centroid within radius) applies symmetrically to both models, so it is not the main issue; neither is the removal of very large lesions in training, since that also applies to both settings. The paper's contribution is plausible and the direction of the result is consistent with the anchor-free literature, but the single untuned baseline prevents full acceptance. The reader's CONDITIONAL verdict remains appropriate; no verdict change is needed.","tokens_in":7025,"tokens_out":3828,"duration_ms":45296,"concrete_test":"Retrain the anchor-based RPN using anchor sizes derived from k-means clustering of the training-set lesion diameters (keeping the same 9-anchor budget: 3 sizes per scale) and evaluate on the same DeepLesion test split. If the tuned anchor-based FROC rises to the anchor-free level (0.735 or above), the reported advantage is an artifact of the hand-picked anchors; if it remains near 0.708 while the anchor-free model stays at 0.735, the central claim is supported. Ideally also repeat both models with at least 3 random seeds and report the mean FROC and range, so the 0.027 gap can be checked against run-to-run variation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the anchor-free RPN outperforms the anchor-based RPN in FROC and speed. Since both models share the same U-Net backbone, detection heads, and loss, the only material difference is the anchor mechanism and the ground-truth assignment. That makes the fairness of the anchor configuration the key assumption. The paper uses exactly one hand-picked anchor set, {3,5,7}, {10,13,17}, {22,30,40} mm, with no sensitivity analysis, no k-means-derived anchors, and no ablation. Anchor-based detectors are well known to be sensitive to anchor scale and aspect ratio, especially for small objects and for objects spanning a wide size range such as DeepLesion lesions (roughly 3 to 70+ mm). Table II further shows the anchor-based RPN drops to 80% sensitivity on lesions >30 mm, where the anchor set stops at 40 mm and only one size is near that band; this is exactly the pattern one would expect from a poorly matched anchor prior. If a better-tuned anchor baseline would reach or exceed 0.735 FROC, the headline 'considerable performance gain' would be an artifact of a weak baseline rather than a genuine property of the anchor-free formulation. No error bars or repeated runs are reported, so tuning luck cannot be distinguished from a real effect. This is an internal comparison risk, not a disagreement with community consensus, and it can be settled empirically.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a 3D anchor-free lesion detector for CT scans, representing each lesion as a center keypoint plus a regressed diameter, and compares it against an anchor-based RPN with the same U-Net/DenseNet backbone. Ground-truth assignment in the anchor-free variant uses positive and non-negative cubes around each lesion centroid, and the classification loss applies a Gaussian penalty to positive samples. Experiments on the DeepLesion dataset report an FROC of 0.735 for the anchor-free RPN versus 0.708 for the anchor-based RPN, with inference times of 1.74 s/scan versus 1.95 s/scan. The paper also reports sensitivity by lesion type and size and argues that the 3D key-slice-agnostic setting is more challenging than prior 2D work.","tokens_in":7245,"tokens_out":5894,"duration_ms":59890,"significance":"If the comparison is valid, the paper would provide a useful data point: anchor-free center-keypoint detection can be applied to 3D lesion detection without loss of accuracy relative to an anchor-based baseline, while simplifying anchor design. The use of a public benchmark (DeepLesion), a shared backbone for both compared models, and a standard detection metric are strengths, and the reported average sensitivities are internally consistent with the table entries. However, the central claim rests on the fairness of the anchor-based baseline and on the reported FROC numbers, and both have unresolved issues: the table reports a metric that does not match the definition in the text, the anchor configuration is hand-picked with no sensitivity analysis, and no uncertainty quantification is given for the small reported gap. The paper is therefore a plausible-but-not-yet-established contribution.","major_comments":[{"comment":"The stated evaluation metric and the reported numbers do not match. Section IV-C defines FROC as the average recall at seven false-positive rates (1/8, 1/4, 1/2, 1, 2, 4, and 8 per scan), but Table I reports six false-positive rates per image (0.5, 1, 2, 4, 8, and 16) and labels their average as \"Avg. FROC\". The reported values 82.17 and 84.01 are averages of those six columns, not of the seven defined rates. Because the central comparison is expressed through this number, the authors must either report the standard metric they define or explicitly justify the modified protocol and explain its relationship to previous work.","section":"§IV-C and Table I"},{"comment":"The fairness of the anchor-based baseline is load-bearing and currently unsupported. Section IV-B states that the anchor-based model uses one hand-picked set of nine anchors ({3,5,7}, {10,13,17}, {22,30,40} mm) with no sensitivity analysis or comparison against alternative configurations. Table II shows that the anchor-based RPN has 80% sensitivity on lesions >30 mm while the anchor-free RPN has 88%, which is the pattern one would expect from an anchor prior that is sparse or poorly matched at the large end of the lesion size distribution. The authors should add an anchor-scale sensitivity study, anchors derived from the training-set size distribution (e.g., k-means), or an ablation over at least one alternative anchor configuration. Without such evidence, the claimed \"considerable performance gain\" of the anchor-free design may be an artifact of an under-tuned baseline.","section":"§IV-B and Table II"},{"comment":"The reported FROC difference is 0.027 (0.735 versus 0.708) and the inference-speed difference is 0.21 s/scan, but the paper reports only a single run and provides no error bars, repeated runs, or statistical significance test. Given that the test set contains 4,912 lesions, this is certainly enough data for a meaningful sensitivity estimate, but a single run does not rule out random initialization or training-procedure variance. The authors should report multiple runs with mean and variance, or an appropriate paired significance test, before claiming a consistent advantage.","section":"§IV-D"},{"comment":"Training and implementation details needed to assess the controlled comparison and to reproduce the results are missing. The paper does not report the optimizer, learning rate schedule, batch size, number of epochs, crop-sampling distribution, NMS threshold, or the IoU threshold used for labeling anchors in the baseline. Since the paper's central claim is a head-to-head comparison between two models that should differ only in the anchor mechanism, these details are necessary to judge whether both models were trained equally well and to allow others to reproduce the comparison.","section":"§IV-B"}],"minor_comments":[{"comment":"The text contains typos: \"this this task\" in Section I, \"grouth truth\" in Section III-B, and \"we the first\" in Section I should be \"we are the first\".","section":"§I and §III-B"},{"comment":"In Eq. (4), the z-component is written as Δcl_z = (cb_z - cp_z)/sl, but the centroid is denoted cg_z elsewhere in the paper; please fix the subscript for consistency.","section":"Equation (4)"},{"comment":"The Gaussian penalty in Eq. (1) is called \"unnormalized\"; please state explicitly whether it is used as a soft weight (not normalized) or clarify the intended normalization, since this affects the interpretation of the loss.","section":"§III-C"},{"comment":"Several symbols in Figure 1 ('up', 'C', 'Dense Block') are not explained in the caption; the strides of the feature maps at each RPN head should also be labeled to match the text.","section":"Figure 1"},{"comment":"The table heading \"FPs per image\" conflicts with the per-scan FROC definition in Section IV-C, and the column \"Avg. FROC\" is not defined; consider renaming it to \"Average sensitivity\" or defining it explicitly.","section":"Table I"},{"comment":"The novelty claim \"we are the first to propose a 3D anchor-free architecture\" is not supported by a thorough related-work search; there is concurrent work on anchor-free 3D detection, and the authors should either cite it or justify why their formulation is distinct.","section":"Related Work"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is early-stage and short, but the core idea is reasonable and the comparison is the right kind of controlled experiment. The main risks are the mismatched FROC definition, the untuned anchor baseline, and the absence of variance estimates. These are fixable with additional experiments, but without them the headline claim is not yet established. I would also ask the authors to carefully verify the novelty claim before final acceptance, since anchor-free 3D detection was emerging at the time of submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a legitimate engineering result, but the size of the claimed gain is uncertain because the anchor-based baseline was never tuned and no error bars are reported. If you take it as 'an anchor-free 3D detector can beat one specific anchor configuration on DeepLesion,' the numbers hold together. If you take it as 'anchor-free is genuinely better,' the paper doesn't yet support that.\n\nWhat's new: they take the CenterNet/CornerNet keypoint idea and extend it to 3D CT for general lesion detection. That's a sensible application. They also show a 3D CNN can work on DeepLesion, where earlier work suggested 2D methods were safer. The FROC numbers are internally consistent (the average matches the 7 false-positive rates), and the inference time comparison is apples-to-apples on the same U-Net backbone, same heads, same loss, differing only in the anchor mechanism and ground-truth assignment. Credit where due: public benchmark, standard metric, and an honest statement that direct comparison with Yan et al. is not possible because of different task settings.\n\nSoft spots, in order of importance. First, the anchor baseline: nine hand-picked sizes, no ablation, no k-means, no sensitivity analysis. Anchor-based detectors are known to be sensitive to this. Table II shows the anchor-based RPN drops to 80% on lesions >30 mm, and the anchor set has only one size near that band (40 mm). That's the pattern you'd expect from a mismatched prior. This is the load-bearing weakness, and it's fixable. Second, no error bars or repeated runs. A 0.027 FROC gap could be tuning luck. Third, training removes lesions >=48 mm (about 11% of data) but testing includes them. That's a train/test mismatch that could bias either model. Fourth, no code or detailed training recipe, which makes reproduction hard. Minor: the epsilon and alpha hyperparameters are hand-set, but they are not fitted to FROC, so that is not circular; it just means a few free parameters are not analyzed.\n\nThe citation pattern is fine. CenterNet and CornerNet are cited, and the 'first 3D anchor-free lesion detector' claim is plausible for this dataset. The related work is adequate.\n\nWho it's for: people building 3D detectors for medical CT, and anyone wanting a data point on anchor-free vs anchor-based in a medical setting. It deserves a serious referee. The right response is to request an anchor ablation, multi-run results, and code or detailed implementation details, not to desk-reject. The central claim is empirical and plausibly correct, just not yet proven at the advertised strength.","headline":"Plausible engineering result, but the anchor-free gain is measured against an untuned anchor baseline and one run, so the size of the claimed advantage isn't established.","tokens_in":7899,"tokens_out":2112,"would_cite":true,"duration_ms":20066,"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 3D anchor-free detector that represents each lesion as a single keypoint outperforms an anchor-based RPN on CT scans in both accuracy and speed.","keywords":["anchor-free object detection","3D lesion detection","computed tomography","keypoint detection","DeepLesion dataset","region proposal network","FROC evaluation","center-point representation"],"falsifier":"Rerun the exact comparison after tuning the anchor-based baseline, for example by deriving anchor sizes from k-means on the DeepLesion diameter distribution or by grid-searching the three per-scale sizes, and check whether the anchor-free FROC gain survives; if a tuned anchor model matches or exceeds 0.735, the central claim collapses.","tokens_in":6739,"feed_emoji":"🩻","tokens_out":5892,"duration_ms":51350,"temperature":0.7,"pith_summary":"This paper tries to establish that the anchor mechanism, widely used in 3D lesion detectors, is unnecessary and even counterproductive when lesions are modeled as single keypoints in CT volumes. It proposes a 3D anchor-free region proposal network in which each lesion is represented by its center point and diameter, and reports better detection accuracy and faster per-scan inference than an anchor-based RPN with the same U-net-style backbone on the DeepLesion dataset. If true, the result would simplify detection pipelines: no anchor templates need to be hand-designed or retuned for a new dataset, and the detector becomes more tolerant of very large lesions. The paper also argues that an anchor-free detector can operate without knowing the key slice in advance, which matters because radiologists do not usually know the key slice before reading a scan.","feed_headline":"A 3D anchor-free detector beats the anchor-based baseline on CT scans","feed_subtitle":"On DeepLesion it lifts FROC from 0.708 to 0.735 and cuts per-scan inference from 1.95s to 1.74s.","key_machinery":"The machinery is a center-keypoint parametrization of a lesion as a 5-element vector {score, coordx, coordy, coordz, diameter}. For each ground-truth box, two concentric cubes are defined: a positive cube of side fraction $\\epsilon_p = 0.8d$ and a non-negative cube of side fraction $\\epsilon_n = 1.2d$. Voxel-center points inside the positive cube are positive training samples, those outside the non-negative cube are negative, and those in between are ignored. Positive centers are weighted by a Gaussian in Euclidean distance to the true centroid ($\\alpha = 1$), classification uses focal loss for negatives plus weighted cross-entropy for positives, and localization regresses stride-normalized center offsets and log diameter with Smooth L1 loss. This assignment replaces anchor-based IoU matching, so no anchor templates or aspect-ratio priors enter the detector.","core_discovery":"The central discovery is comparative: on DeepLesion, replacing anchor-based proposal generation with a center-keypoint formulation improves the FROC metric from 0.708 to 0.735 and reduces per-scan inference time from 1.95s to 1.74s, with the largest gains at low false-positive rates, for example sensitivity at 0.5 false positives per image rises from 65.74% to 68.73%. The paper interprets this as evidence that overlapped anchors are unnecessary for lesions because lesions in CT scans do not overlap, and that the anchor-free formulation is more tolerant of large lesions. It positions the contribution as the first 3D anchor-free architecture for general lesion detection.","pith_inferences":["If the advantage holds across datasets, the anchor-free design could remove an entire hyperparameter-tuning step from medical object detection: the same detector could be dropped onto new scan types without re-deriving anchor dimensions from bounding-box statistics.","The non-overlap assumption is the likely boundary of the method; for organs or tumors that do overlap, or for very elongated structures, a single center keypoint may lose information that anchor templates carry, so the advantage may not transfer outside roughly spherical lesions.","A direct test of the tolerance claim would be to train without removing lesions of diameter at least 48 mm and compare training stability; the paper's decision to exclude them during training leaves open whether the apparent tolerance to large lesions is architectural or an artifact of that exclusion."],"forward_implications":["A 3D lesion detector can be built without any hand-designed anchor sizes; the anchor-free head simply predicts a center keypoint and diameter.","Removing anchors cuts false-positive pressure and speeds up inference because fewer candidate boxes need to be suppressed.","The center-keypoint formulation is more tolerant of large lesions, those with diameter above 30 mm, where anchor-based matching loses sensitivity.","Detection on a 60-mm CT chunk succeeds without knowing the key slice in advance, unlike approaches that depend on the key slice being given.","The same U-net backbone yields higher FROC at low false-positive rates when trained with center-keypoint assignment rather than IoU-based anchor assignment."],"supporting_citations":[{"why":"Supplies all training, validation, and test CT scans and lesion annotations for the experiments.","marker":"[20]"},{"why":"Supplies the center-keypoint anchor-free formulation that the detector adapts to 3D.","marker":"[21]"},{"why":"Supplies the Gaussian penalization of positive center points used in the classification loss.","marker":"[22]"},{"why":"Supplies the focal loss used for negative samples in the classification branch.","marker":"[23]"},{"why":"Defines the FROC evaluation protocol with the seven false-positive rates used to compute the reported average recall.","marker":"[4]"},{"why":"Supplies the dense blocks that compose the U-net-style backbone of both the anchor-based and anchor-free detectors.","marker":"[6]"}],"fun_headline_variants":["Anchor-free 3D detector beats anchors on CT lesions","Keypoint-based CT lesion detector outruns anchor baseline","3D anchor-free lesion detection: faster and more accurate","Drop anchors to boost CT lesion detection in 3D"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes the hand-picked anchor sizes for the baseline are a fair, reasonably tuned configuration, with no reported ablation or sensitivity analysis, so the measured gains could be an artifact of an under-tuned baseline.","fun_headline_variants_meta":{"raw":{"variants":["Anchor-free 3D detector beats anchors on CT lesions","Keypoint-based CT lesion detector outruns anchor baseline","3D anchor-free lesion detection: faster and more accurate","Drop anchors to boost CT lesion detection in 3D"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000203,"raw_usage":{"total_tokens":1324,"prompt_tokens":825,"completion_tokens":499,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":441,"completion_tokens_details":{"reasoning_tokens":432}},"tokens_in":441,"tokens_out":499,"duration_ms":4951,"temperature":1.0,"reasoning_tokens":432,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:16:56.397794+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the exact comparison after tuning the anchor-based baseline, for example by deriving anchor sizes from k-means on the DeepLesion diameter distribution or by grid-searching the three per-scale sizes, and check whether the anchor-free FROC gain survives; if a tuned anchor model matches or exceeds 0.735, the central claim collapses.","supporting_citations":[{"cited_title":"Deep lesion graphs in the wild: Relationship learning and organization of significant radiology image findings in a diverse large- scale lesion database,","cited_arxiv_id":null,"evidence_quote":"Supplies all training, validation, and test CT scans and lesion annotations for the experiments."},{"cited_title":"Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in computed tomography images: the luna16 challenge,","cited_arxiv_id":null,"evidence_quote":"Defines the FROC evaluation protocol with the seven false-positive rates used to compute the reported average recall."},{"cited_title":"Densely connected convolutional networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the dense blocks that compose the U-net-style backbone of both the anchor-based and anchor-free detectors."}],"review_version":1}