Pith. sign in

REVIEW 4 major objections 6 minor 23 references

3D Anchor-Free Lesion Detector on Computed Tomography Scans

T0 review · 4 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read 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.

desk verdict 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. read the letter →

arxiv 1908.11324 v1 pith:RKAOOLSV submitted 2019-08-29 cs.CV

Ning Zhang , Dechun Wang , Xinzi Sun , Pengfei Zhang , Chenxi Zhang , Yu Cao , Benyuan Liu This is my paper · ORCID
classification cs.CV
keywords anchor-freeobjectdetection3DlesioncomputedtomographykeypointDeepdatasetregionproposalnetworkFROCevaluationcenter-pointrepresentation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

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.

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 (4)
  1. [§IV-C and Table I] 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.
  2. [§IV-B and Table II] 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.
  3. [§IV-D] 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.
  4. [§IV-B] 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.
minor comments (6)
  1. [§I and §III-B] 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".
  2. [Equation (4)] 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.
  3. [§III-C] 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.
  4. [Figure 1] 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.
  5. [Table I] 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.
  6. [Related Work] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the anchor-free/anchor-based comparison is an empirical evaluation on held-out DeepLesion data.

full rationale

The paper's central claim is an empirical comparison on the held-out DeepLesion test set. The anchor-free RPN and the anchor-based RPN share the same U-Net/DenseNet backbone, detection heads, and loss formulation, and they differ only in the anchor mechanism and ground-truth assignment. The reported FROC scores (0.735 vs 0.708) and inference times (1.74s vs 1.95s) are measured quantities, not quantities derived from the model definitions. The Gaussian penalty weight in Eq. (1) and the positive/non-negative cube fractions (epsilon_p = 0.8, epsilon_n = 1.2) are fixed hyperparameters chosen before evaluation; they are not fitted to the target FROC metric, so no fitted input is being renamed as a prediction. The localization offsets in Eq. (4) are direct regression targets with stride normalization, not a re-encoding of the anchor-based outputs. The hand-picked anchor sizes in Section IV-B could make the anchor-based baseline weak, but that is an experimental-fairness or correctness concern, not circularity: the anchor-free numbers are not constructed from the anchor-based numbers. Self-citations in the reference list ([7]-[12]) appear only as unrelated prior work by the same group and are not load-bearing for any claim. No step in the derivation chain reduces to its own input by definition, and no prediction is statistically forced by a fitted parameter. The comparison is self-contained against an external benchmark with an external evaluation metric, so no significant circularity is present.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The central claim depends on: (a) two hand-set epsilon values defining the assignment regions, (b) a hand-set Gaussian alpha, (c) a hand-picked anchor configuration for the baseline, (d) a post-hoc training exclusion of large lesions, and (e) the unvalidated 2D-to-3D annotation conversion. No new physical entities are introduced. The number of free parameters is typical for a deep learning paper, but the lack of ablations makes it hard to know which choices drive the reported gain.

free parameters (5)
  • positive cube epsilon (epsilon_p) = 0.8
    Ground truth assignment uses epsilon_p=0.8 to define the positive cube. Chosen by hand (Section III-B), no ablation reported.
  • non-negative cube epsilon (epsilon_n) = 1.2
    Ground truth assignment uses epsilon_n=1.2 for the ignored region (Section III-B). Chosen by hand, no ablation.
  • Gaussian penalty alpha = 1
    Eq. (1) uses alpha=1 to down-weight positive points far from the centroid. Not tuned, following [22].
  • Anchor sizes (3 per scale) = {3,5,7}, {10,13,17}, {22,30,40}
    Anchor configuration for the baseline is hand-picked (Section IV-B). No tuning or sensitivity analysis, directly determines the baseline quality.
  • Lesion size threshold for training exclusion = 48mm
    Lesions >=48mm (~11% of training data) are removed during training (Section IV-B) because they cause convergence issues. This is a post-hoc exclusion that affects the comparison.
assumptions (3)
  • domain assumption DeepLesion annotations converted to 3D form {X, Y, Z, Diameter} from 2D key-slice annotations are correct enough for training a 3D detector.
    Section IV-A: 'The 2D annotations are approximately converted to 3D ones with the form of {X, Y, Z, Diameter}.' The accuracy of this conversion directly affects all reported numbers, but no validation of the conversion is provided.
  • domain assumption FROC metric with the paper's True Positive criterion (centroid distance less than ground-truth radius) is a valid measure of detection performance.
    Section IV-C defines TP based on centroid distance, which is a looser criterion than IoU. This is inherited from LUNA16 but is an assumption about clinical relevance.
  • standard math The U-Net/DenseNet backbone and the described training procedure are sufficient to learn a good keypoint detector; no theoretical guarantee is provided.
    The paper relies on standard deep learning practice (architectures, focal loss, smooth L1). These are not proved in the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of 3D Anchor-Free Lesion Detector on Computed Tomography Scans." pith.science (2026). https://pith.science/paper/RKAOOLSV

@misc{pith2026190811324,
  author       = {Pith},
  title        = {Pith review of: 3D Anchor-Free Lesion Detector on Computed Tomography Scans},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RKAOOLSV}},
  note         = {Machine review of arXiv:1908.11324}
}
read the original abstract

Lesions are injuries and abnormal tissues in the human body. Detecting lesions in 3D Computed Tomography (CT) scans can be time-consuming even for very experienced physicians and radiologists. In recent years, CNN based lesion detectors have demonstrated huge potentials. Most of current state-of-the-art lesion detectors employ anchors to enumerate all possible bounding boxes with respect to the dataset in process. This anchor mechanism greatly improves the detection performance while also constraining the generalization ability of detectors. In this paper, we propose an anchor-free lesion detector. The anchor mechanism is removed and lesions are formalized as single keypoints. By doing so, we witness a considerable performance gain in terms of both accuracy and inference speed compared with the anchor-based baseline

Figures

Figures reproduced from arXiv: 1908.11324 by the authors.

Figure 1
Figure 1. The architecture of the whole network. The network is of a “U” structure and consists of an upstream and a downstream pathway. Upstream and [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Visualization of lesion types: bone, kidney and soft tissue. Red and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 17 canonical work pages

  1. [1]

    Ct screening for lung cancer: frequency and significance of part-solid and nonsolid nodules,

    C. I. Henschke, D. F. Yankelevitz, R. Mirtcheva, G. McGuinness, D. Mc- Cauley, and O. S. Miettinen, “Ct screening for lung cancer: frequency and significance of part-solid and nonsolid nodules,” American Journal of Roentgenology , vol. 178, no. 5, pp. 1053–1057, 2002

    work page 2002

  2. [2]

    A large-scale evaluation of automatic pulmonary nodule detection in chest ct using local image features and k-nearest-neighbour classification,

    K. Murphy, B. van Ginneken, A. M. Schilham, B. De Hoop, H. Gietema, and M. Prokop, “A large-scale evaluation of automatic pulmonary nodule detection in chest ct using local image features and k-nearest-neighbour classification,” Medical image analysis , vol. 13, no. 5, pp. 757–770, 2009

    work page 2009

  3. [3]

    Automatic detection of subsolid pulmonary nodules in thoracic computed tomography images,

    C. Jacobs, E. M. van Rikxoort, T. Twellmann, E. T. Scholten, P. A. de Jong, J.-M. Kuhnigk, M. Oudkerk, H. J. de Koning, M. Prokop, C. Schaefer-Prokop et al. , “Automatic detection of subsolid pulmonary nodules in thoracic computed tomography images,” Medical image analysis, vol. 18, no. 2, pp. 374–384, 2014

    work page 2014

  4. [4]

    Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in computed tomography images: the luna16 challenge,

    A. A. A. Setio, A. Traverso, T. De Bel, M. S. Berens, C. van den Bogaard, P. Cerello, H. Chen, Q. Dou, M. E. Fantacci, B. Geurts et al., “Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in computed tomography images: the luna16 challenge,” Medical image analysis , vol. 42, pp. 1–13, 2017

    work page 2017

  5. [5]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 770–778

    2016

  6. [6]

    Densely connected convolutional networks,

    G. Huang, Z. Liu, K. Q. Weinberger, and L. van der Maaten, “Densely connected convolutional networks,” in Proceedings of the IEEE confer- ence on computer vision and pattern recognition , vol. 1, no. 2, 2017, p. 3

    work page 2017

  7. [7]

    Improved multimodal represen- tation learning with skip connections,

    N. Zhang, Y . Cao, B. Liu, and Y . Luo, “Improved multimodal represen- tation learning with skip connections,” in Proceedings of the 2017 ACM on Multimedia Conference , ser. MM ’17. New York, NY , USA: ACM, 2017, pp. 654–662

    work page 2017

  8. [8]

    Improving tuberculosis diagnostics using deep learning and mobile health tech- nologies among resource-poor and marginalized communities,

    Y . Cao, C. Liu, B. Liu, M. J. Brunette, N. Zhang, T. Sun, P. Zhang, J. Peinado, E. S. Garavito, L. L. Garcia, and W. H. Curioso, “Improving tuberculosis diagnostics using deep learning and mobile health tech- nologies among resource-poor and marginalized communities,” in 2016 IEEE First International Conference on Connected Health: Applications, Systems ...

    work page 2016

Show all 23 references
  1. [9]

    ihear food: Eating detection using commodity bluetooth headsets,

    Y . Gao, N. Zhang, H. Wang, X. Ding, X. Ye, G. Chen, and Y . Cao, “ihear food: Eating detection using commodity bluetooth headsets,” in 2016 IEEE First International Conference on Connected Health: Applications, Systems and Engineering Technologies (CHASE) , June 2016

    2016

  2. [10]

    Improving tuberculosis diagnostics using deep learning and mobile health technologies among resource- poor communities in per ´u,

    M. F. Alcantara, Y . Cao, C. Liu, B. Liu, M. Brunette, N. Zhang, T. Sun, P. Zhang, Q. Chen, Y . Li et al. , “Improving tuberculosis diagnostics using deep learning and mobile health technologies among resource- poor communities in per ´u,” Smart Health , vol. 1, pp. 66–76, 2017

    2017

  3. [11]

    Heterospark: A heterogeneous cpu/gpu spark platform for machine learning algorithms,

    P. Li, Y . Luo, N. Zhang, and Y . Cao, “Heterospark: A heterogeneous cpu/gpu spark platform for machine learning algorithms,” in 2015 IEEE International Conference on Networking, Architecture and Storage (NAS), Aug 2015, pp. 347–348

    2015

  4. [12]

    People re-identification by multi-branch cnn with multi-scale features,

    X. Sun, N. Zhang, Q. Chen, Y . Cao, and B. Liu, “People re-identification by multi-branch cnn with multi-scale features,” in 2019 26th IEEE International Conference on Image Processing (ICIP) . IEEE, 2019

    2019

  5. [13]

    3d context enhanced region- based convolutional neural network for end-to-end lesion detection,

    K. Yan, M. Bagheri, and R. M. Summers, “3d context enhanced region- based convolutional neural network for end-to-end lesion detection,” in International Conference on Medical Image Computing and Computer- Assisted Intervention . Springer, 2018, pp. 511–519

    2018

  6. [14]

    Towards single-phase single-stage detection of pulmonary nodules in chest ct imaging,

    Z. Xie, “Towards single-phase single-stage detection of pulmonary nodules in chest ct imaging,” arXiv preprint arXiv:1807.05972 , 2018

    2018 arXiv

  7. [15]

    Evaluate the malignancy of pulmonary nodules using the 3d deep leaky noisy-or network,

    F. Liao, M. Liang, Z. Li, X. Hu, and S. Song, “Evaluate the malignancy of pulmonary nodules using the 3d deep leaky noisy-or network,” arXiv preprint arXiv:1711.08324, 2017

    2017 arXiv

  8. [16]

    A closer look: Small object detection in faster r-cnn,

    C. Eggert, S. Brehm, A. Winschel, D. Zecha, and R. Lienhart, “A closer look: Small object detection in faster r-cnn,” in Multimedia and Expo (ICME), 2017 IEEE International Conference on . IEEE, 2017, pp. 421–426

    2017

  9. [17]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in International Conference on Medical image computing and computer-assisted intervention . Springer, 2015, pp. 234–241

    2015

  10. [18]

    Feature pyramid networks for object detection,

    T.-Y . Lin, P. Doll´ar, R. Girshick, K. He, B. Hariharan, and S. Belongie, “Feature pyramid networks for object detection,” in CVPR, vol. 1, no. 2, 2017, p. 4

    2017

  11. [19]

    Deeplung: Deep 3d dual path nets for automated pulmonary nodule detection and classification,

    W. Zhu, C. Liu, W. Fan, and X. Xie, “Deeplung: Deep 3d dual path nets for automated pulmonary nodule detection and classification,” arXiv preprint arXiv:1801.09555, 2018

    2018 arXiv

  12. [20]

    Deep lesion graphs in the wild: Relationship learning and organization of significant radiology image findings in a diverse large- scale lesion database,

    K. Yan, X. Wang, L. Lu, L. Zhang, A. P. Harrison, M. Bagheri, and R. M. Summers, “Deep lesion graphs in the wild: Relationship learning and organization of significant radiology image findings in a diverse large- scale lesion database,” in The IEEE Conference on Computer Vision ...

    2018

  13. [21]

    Objects as points,

    X. Zhou, D. Wang, and P. Kr ¨ahenb¨uhl, “Objects as points,” CoRR, vol. abs/1904.07850, 2019. [Online]. Available: http://arxiv.org/abs/1904. 07850

    1904 arXiv

  14. [22]

    Cornernet: Detecting objects as paired keypoints,

    H. Law and J. Deng, “Cornernet: Detecting objects as paired keypoints,” CoRR, vol. abs/1808.01244, 2018. [Online]. Available: http://arxiv.org/abs/1808.01244

    2018 arXiv

  15. [23]

    Focal loss for dense object detection,

    T.-Y . Lin, P. Goyal, R. Girshick, K. He, and P. Doll ´ar, “Focal loss for dense object detection,” arXiv preprint arXiv:1708.02002 , 2017

    2017 arXiv

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.