Pith. sign in

REVIEW 4 major objections 5 minor 31 references

PDSE: A Multiple Lesion Detector for CT Images using PANet and Deformable Squeeze-and-Excitation Block

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The authors propose PDSE, a one-stage RetinaNet-based detector combining path aggregation and deformable squeeze-and-excitation attention, and report a mean average precision of 0.2053 on the DeepLesion benchmark, which they claim is a…

desk verdict A plausible module combination undermined by a missing SOTA comparison; the central performance claim is unsupported. read the letter →

arxiv 2506.03608 v1 pith:JDUAHN4H submitted 2025-06-04 cs.CV

Di Fan , Heng Yu , Zhiyuan Xu This is my paper
classification cs.CV
keywords CTlesiondetectionRetinaNetPANetDeformableconvolutionSqueeze-and-ExcitationDeepbenchmarkone-stagedetectorsmallobject
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 a one-stage detector can outperform the two-stage and anchor-free detectors currently used for multi-organ lesion detection in CT scans. The proposed PDSE model starts from RetinaNet and adds two components: a path aggregation network that feeds low-level feature maps into the pyramid, and a Deformable Squeeze-and-Excitation (DSE) block that combines deformable convolution with channel attention. On the DeepLesion dataset the authors report a mean average precision of 0.2053, above their own RetinaNet baseline and above Faster R-CNN, CenterNet, and FSAF under the same evaluation protocol. The claim is that this is a new state-of-the-art result for detectors that take 2D CT images as input.

What carries the argument

The Deformable Squeeze-and-Excitation (DSE) block is the central new component. It couples a deformable convolution, which shifts each sampling point by a learned 2D offset to follow lesion geometry, with a squeeze-and-excitation attention branch that compresses each channel to a scalar, learns per-channel weights, and reweights the feature map; a local attention branch runs in parallel with the global SE branch. The block is inserted between feature pyramid levels, while PANet supplies low-level feature maps through the pyramid. This machinery is what the paper credits for improved detection of small and multiscaled lesions.

What would settle it

Run 3DCE, MULAN, and ULDOR on the same 70/15/15 DeepLesion split and the same mAP evaluation; if any of them scores above 0.2053, the paper's new-SOTA claim is refuted. Alternatively, if the official DeepLesion test split gives PDSE a materially lower mAP than 0.2053, the reported number would not transfer.

Watch

Extended reading notes

Core claim

The paper's central claim is that the combination of PANet-style path aggregation and a Deformable Squeeze-and-Excitation block turns RetinaNet into a universal lesion detector that reaches 0.2053 mAP on DeepLesion's nine lesion categories. The DSE block adapts the convolutional sampling grid to the irregular shape of lesions and reweights feature channels so local details are fused into a global context. The authors report that the largest per-class gains occur for kidney and lung lesions, while abdominal lesions remain the hardest category. They conclude that the model outperforms other detectors for 2D image input on this dataset.

Load-bearing premise

The claim of a new state of the art depends on the assumption that the authors' random 70/15/15 split of DeepLesion and their own re-implementations of comparison detectors give numbers that can be compared directly with published results of 3DCE, MULAN, and ULDOR, which the paper does not re-run.

Editorial extensions

If this is right

  • PDSE reports 0.2053 mAP on DeepLesion, higher than its own RetinaNet and RetinaNet+PANet baselines under the same protocol.
  • The largest per-class gains are in kidney and lung lesions, while the abdomen category remains the most difficult, indicating where future effort is needed.
  • The paper's comparison indicates that anchor-free detectors (CenterNet, FSAF) lag behind anchor-based RetinaNet variants on this dataset.
  • The one-stage design avoids a region proposal network, which the paper argues makes lesion detection faster and more accurate than two-stage alternatives.

Reading between the lines

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

  • The DSE block is architecture-agnostic and could be tested on natural-image object detection to see whether the gain is specific to CT lesions or generalizes.
  • DeepLesion annotations include RECIST diameters, so a multi-task head predicting lesion size from the same features is a natural next step that the paper does not attempt.
  • The per-class pattern suggests a focused study on small, low-contrast lesions would reveal where the attention mechanism contributes most.
  • The split and preprocessing used here differ across DeepLesion papers; a shared evaluation protocol would let the 0.2053 mAP be compared directly with published results of 3DCE, MULAN, and ULDOR.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes PDSE, a one-stage lesion detector for CT images built by combining RetinaNet with a Path Aggregation Network (PANet) and a Deformable Squeeze-and-Excitation (DSE) block. On the DeepLesion benchmark, the authors report an mAP of 0.2053 and claim this constitutes new state-of-the-art performance. The paper presents per-class AP values for PDSE and for several self-implemented baselines (RetinaNet, RetinaNet+PANet, Faster R-CNN, CenterNet, FSAF), an ablation, and a qualitative example.

Significance. If the central claim were substantiated, the paper would show that a one-stage RetinaNet/PANet/DSE hybrid surpasses prior universal lesion detectors on DeepLesion. However, the claim is not supported by the evidence as presented: the cited state-of-the-art methods (3DCE, MULAN, ULDOR) are absent from the quantitative comparison, the evaluation protocol is underspecified, and the reported mAP of 0.2053 is below the roughly 0.30-0.33 mAP values that the cited methods report on this dataset. The paper's strength is its ablation isolating the contributions of PANet and DSE, and its use of the public DeepLesion benchmark, but the novelty is an incremental combination of established components. No code or trained models are released, so reproducibility is limited.

major comments (4)
  1. [§3.2, Table 1] The headline claim of "new state-of-the-art performance" is not supported by the comparison in Table 1. The table includes only RetinaNet, RetinaNet+PANet, Ours, Faster R-CNN, CenterNet, and FSAF. The methods cited in the Introduction as state of the art (3DCE, MULAN, ULDOR) are not evaluated quantitatively; ULDOR is listed in the method list in §3.2 but appears only in the qualitative Figure 3. The sentence "RetinaNet achieves a higher mAP than Faster R-CNN and ULDOR" is therefore not verifiable for ULDOR. Because no published results are reproduced or cited, Table 1 does not establish superiority over existing detectors.
  2. [§3.4, Table 1] The overall mAP values cited in the text, 0.2053 for Ours and 0.0718 for CenterNet, cannot be verified from Table 1 because the table has no overall mAP row. Moreover, the per-class numbers do not consistently favor the proposed method: RetinaNet is higher than Ours on Abdomen (0.0974 vs. 0.0963), Lung (0.2466 vs. 0.2456), and Pelvis (0.1121 vs. 0.1066). The claim that the method "significantly improves" detection over the baselines needs error bars or significance tests, especially given the small per-class differences.
  3. [§3.1] The evaluation protocol is underspecified. The paper states only that DeepLesion is split 70/15/15; it does not specify whether the split is patient-level, random, or the standard split used by prior DeepLesion studies, nor does it give the seed or the number of test images. The mAP definition is also missing: no IoU threshold, matching rule, or averaging convention is given. Without this information, the reported numbers cannot be compared with published results, and the claim that 0.2053 is state of the art is untestable.
  4. [§2.2] The DSE block is described only by a block diagram and qualitative prose. There are no equations defining the deformable offsets, the squeeze/excitation/reweight operations, the combination of global and local attention, or the channel dimensions and reduction ratios. This makes the architectural contribution difficult to reproduce and makes it unclear whether DSE is a simple composition of deformable convolution followed by an SE block or a genuinely new module.
minor comments (5)
  1. [§3.2, Table 1] The text says "we present the mAP results for eight specific types of lesions," but Table 1 has nine rows (Bone, Abdomen, Mediastinum, Liver, Lung, Kidney, Tissue, Pelvis, Other). The column header calls the table "mAP for different algorithms comparison," although the entries are per-class AP values, not mAP.
  2. [§2, Figure 1] The dimensions in Figure 1 are inconsistent: P2^PA is listed as (128,128,256), but the corresponding DSE/class subnet entry is labeled (64,64,512). Either the figure is mislabeled or the deformable module halves the spatial resolution, which is not explained in the text.
  3. [§3.3, Figure 3] The qualitative comparison is a single example slice. The caption describes colors for the different methods, but the legend of the figure is ambiguous and does not clearly map the colored boxes to the methods, especially for ULDOR.
  4. [Throughout] There are multiple typos and wording issues, including "RestinaNet" in the keywords, "Path Aggragation Net" in §2, "pramid" in §2, and inconsistent capitalization of PANet/PAnet. The manuscript would benefit from a careful proofread.
  5. [§2.1] The description of PANet cites instance segmentation, but the paper does not explain how adaptive feature pooling is adapted for the one-stage RetinaNet framework, nor does it clarify which parts of PANet are used (bottom-up path augmentation, adaptive pooling, or both).

Circularity Check

0 steps flagged · score 0.0 of 10

No derivational circularity: PDSE's mAP is a measured empirical outcome, and the self-citations are not load-bearing; the unsupported SOTA comparison is an evaluation-protocol weakness, not circularity.

full rationale

The paper makes no formal derivation that could reduce to its own inputs. PDSE is a RetinaNet + PANet + deformable SE architecture trained and tested on a 70/15/15 split of DeepLesion, so the reported mAP 0.2053 is an empirically measured result rather than a fitted parameter renamed as a prediction. The authors' self-citations (Refs. 21 and 29) are used only as general design and related-work references; removing them would not change the architecture, training procedure, or reported numbers, so they are not load-bearing. The headline claim of 'new state-of-the-art performance' is not supported because Section 3.2 lists ULDOR as a comparison method but Table 1 omits ULDOR, 3DCE, and MULAN, and no published SOTA numbers are reproduced under a matched protocol. That is a comparison-protocol and correctness weakness, not circularity: the claim does not become true by definition, and the result is not entailed by the table's comparison set. No equation, fitted parameter, imported uniqueness theorem, or self-citation chain makes the output equivalent to an input, so no circular step can be exhibited.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The paper is an empirical machine learning study; no new theoretical entities or forces are introduced. The free parameters are the standard hyperparameters of the model, which are not fully reported, making the central claim dependent on undisclosed choices. The axioms are domain assumptions about the evaluation protocol.

free parameters (4)
  • RetinaNet anchor configuration = not reported
    Anchor scales and aspect ratios are not specified; they are crucial for detection performance and are typically hand-tuned or set by default.
  • Focal loss hyperparameters = not reported
    No values given; these directly affect classification of hard examples.
  • DSE reduction ratio and deformable group settings = not reported
    The architecture figure indicates FC layers but no channel reduction ratio; the number of deformable convolution offsets is unspecified.
  • Input image resolution = not reported
    The figure shows C2 as 256x256 but the dataset states 512x512; preprocessing or resizing is not described.
assumptions (3)
  • domain assumption The random 70/15/15 split of DeepLesion produces a test set comparable to the official benchmark split.
    Section 3.1 states the split; no justification that it matches the standard protocol used by compared methods.
  • domain assumption mAP is computed with a standard object detection metric (e.g., PASCAL VOC IoU=0.5), as used by the cited methods.
    No metric details are provided, yet validity of comparing mAP values depends on identical evaluation protocol.
  • domain assumption The HU clipping range [-1024, 3071] preserves all relevant soft-tissue information.
    Section 3.1; standard practice, but no ablation is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PDSE: A Multiple Lesion Detector for CT Images using PANet and Deformable Squeeze-and-Excitation Block." pith.science (2026). https://pith.science/paper/JDUAHN4H

@misc{pith2026250603608,
  author       = {Pith},
  title        = {Pith review of: PDSE: A Multiple Lesion Detector for CT Images using PANet and Deformable Squeeze-and-Excitation Block},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JDUAHN4H}},
  note         = {Machine review of arXiv:2506.03608}
}
read the original abstract

Detecting lesions in Computed Tomography (CT) scans is a challenging task in medical image processing due to the diverse types, sizes, and locations of lesions. Recently, various one-stage and two-stage framework networks have been developed to focus on lesion localization. We introduce a one-stage lesion detection framework, PDSE, by redesigning Retinanet to achieve higher accuracy and efficiency for detecting lesions in multimodal CT images. Specifically, we enhance the path aggregation flow by incorporating a low-level feature map. Additionally, to improve model representation, we utilize the adaptive Squeeze-and-Excitation (SE) block and integrate channel feature map attention. This approach has resulted in achieving new state-of-the-art performance. Our method significantly improves the detection of small and multiscaled objects. When evaluated against other advanced algorithms on the public DeepLesion benchmark, our algorithm achieved an mAP of over 0.20.

Figures

Figures reproduced from arXiv: 2506.03608 by the authors.

Figure 1
Figure 1. Framework structure for PDSE, where PA is the Path Aggragation Net (PANet) and deformSE is Deformable Squeeze-and-Excitation (DSE) block. third column networks with PANet, the third column network and the fourth column networks with deformable SE blocks. We first describe how path aggre￾gation improves RetinaNet. We then show the mechanism for further improving detection performance with Deformable Squeeze-and-Excit… view at source ↗
Figure 2
Figure 2. Framework for PDSE. In [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. illustrates the nine types of lesions and the qualitative lesion detection results produced by the models trained with the configurations mentioned above. In the figure, the red box denotes the ground truth, the green box represents RetinaNet, the blue box represents RetinaNet + PANet, the black box represents RetinaNet + PANet + DSE, and the purple box represents ULDOR. From the figure, it is evident that our metho… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 25 canonical work pages

  1. [1]

    In: Deep learning and data labeling for medical applications, pp

    Ben-Cohen, A., Diamant, I., Klang, E., Amitai, M., Greenspan, H.: Fully convolu- tional network for liver segmentation and lesions detection. In: Deep learning and data labeling for medical applications, pp. 77–85. Springer (2016)

    work page 2016

  2. [2]

    In: Medical Imaging 2024: Image-Guided Procedures, Robotic Interventions, and Modeling

    Chang, Q., Daneshpajooh, V., Byrnes, P.D., Ahmad, D., Toth, J., Bascom, R., Higgins, W.E.: Synchronization and analysis of multimodal bronchoscopic airway exams for early lung cancer detection. In: Medical Imaging 2024: Image-Guided Procedures, Robotic Interventions, and Modeling. vol. 12928, pp. 283–290. SPIE (2024)

    work page 2024

  3. [3]

    In: Proceedings of the IEEE international conference on computer vision

    Dai, J., Qi, H., Xiong, Y., Li, Y., Zhang, G., Hu, H., Wei, Y.: Deformable convolu- tional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 764–773 (2017)

    2017

  4. [4]

    In: Proceedings of the IEEE International Conference on Com- puter Vision

    Duan, K., Bai, S., Xie, L., Qi, H., Huang, Q., Tian, Q.: Centernet: Keypoint triplets for object detection. In: Proceedings of the IEEE International Conference on Com- puter Vision. pp. 6569–6578 (2019)

    work page 2019

  5. [5]

    In: Proceedings of the IEEE international conference on computer vision

    He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 2961–2969 (2017)

    2017

  6. [6]

    He,K.,Zhang,X.,Ren,S.,Sun,J.:Deepresiduallearningforimagerecognition.In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770–778 (2016) Title Suppressed Due to Excessive Length 9

    work page 2016

  7. [7]

    In: 2020 IEEE 17th International Symposium on Biomedical Imaging (ISBI)

    Henry, H.Y., Feng, X., Wang, Z., Sun, H.: Mixmodule: Mixed cnn kernel module for medical image segmentation. In: 2020 IEEE 17th International Symposium on Biomedical Imaging (ISBI). pp. 1508–1512. IEEE (2020)

    work page 2020

  8. [8]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Hu, J., Shen, L., Sun, G.: Squeeze-and-excitation networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7132–7141 (2018)

    2018

Show all 31 references
  1. [9]

    In: 2017 IEEE 14th International Symposium on Biomed- ical Imaging (ISBI 2017)

    Huang, X., Shan, J., Vaidya, V.: Lung nodule detection in ct using 3d convolu- tional neural networks. In: 2017 IEEE 14th International Symposium on Biomed- ical Imaging (ISBI 2017). pp. 379–383. IEEE (2017)

    2017

  2. [10]

    Magnetic Resonance in Medical Sciences 23(2), 184–192 (2024)

    Iwamura, M., Ide, S., Sato, K., Kakuta, A., Tatsuo, S., Nozaki, A., Wakayama, T., Ueno, T., Haga, R., Kakizaki, M., et al.: Thin-slice two-dimensional t2-weighted imaging with deep learning-based reconstruction: improved lesion detection in the brain of patients with multiple ...

    2024

  3. [11]

    Computers in Biology and Medicine 157, 106726 (2023)

    Jiang, H., Diao, Z., Shi, T., Zhou, Y., Wang, F., Hu, W., Zhu, X., Luo, S., Tong, G., Yao, Y.D.: A review of deep learning-based multiple-lesion recognition from medical images: classification, detection and segmentation. Computers in Biology and Medicine 157, 106726 (2023)

    2023

  4. [12]

    Annals of surgery274(6), e1153–e1161 (2021)

    Jiang, Y., Jin, C., Yu, H., Wu, J., Chen, C., Yuan, Q., Huang, W., Hu, Y., Xu, Y., Zhou, Z., et al.: Development and validation of a deep learning ct signature to predict survival and chemotherapy benefit in gastric cancer: a multicenter, retro- spective study. Annals of surge...

    2021

  5. [13]

    Nature communications12(1), 1851 (2021)

    Jin, C., Yu, H., Ke, J., Ding, P., Yi, Y., Jiang, X., Duan, X., Tang, J., Chang, D.T., Wu, X., et al.: Predicting treatment response from longitudinal images using multi-task deep learning. Nature communications12(1), 1851 (2021)

    2021

  6. [14]

    The International Archives of Pho- togrammetry, Remote Sensing and Spatial Information Sciences42, 13 (2017)

    Kalinovsky, A., Liauchuk, V., Tarasau, A.: Lesion detection in ct images using deep learning semantic segmentation technique. The International Archives of Pho- togrammetry, Remote Sensing and Spatial Information Sciences42, 13 (2017)

    2017

  7. [15]

    Radiology310(1), e230981 (2024)

    van Leeuwen, K.G., Schalekamp, S., Rutten, M.J., Huisman, M., Schaefer-Prokop, C.M., de Rooij, M., van Ginneken, B., Maresch, B., Geurts, B.H., van Dijke, C.F., et al.: Comparison of commercial ai software performance for radiograph lung nodule detection and bone age predictio...

    2024

  8. [16]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Lin, T.Y., Dollár, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2117–2125 (2017)

    2017

  9. [17]

    Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollár, P.: Focal loss for dense object detection.In:ProceedingsoftheIEEEinternationalconferenceoncomputervision. pp. 2980–2988 (2017)

    2017

  10. [18]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    Liu, S., Qi, L., Qin, H., Shi, J., Jia, J.: Path aggregation network for instance segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8759–8768 (2018)

    2018

  11. [19]

    Pawlowski, N., Lee, M.C., Rajchl, M., McDonagh, S., Ferrante, E., Kamnitsas, K., Cooke, S., Stevenson, S., Khetani, A., Newman, T., et al.: Unsupervised lesion detection in brain ct using bayesian convolutional autoencoders (2018)

    2018

  12. [20]

    In: Advances in neural information processing systems

    Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detec- tion with region proposal networks. In: Advances in neural information processing systems. pp. 91–99 (2015)

    2015

  13. [21]

    Medical Physics 51(4), 2733–2740 (2024)

    Song, W., Yu, H., Wu, J.: Plu-net: extraction of multiscale feature fusion. Medical Physics 51(4), 2733–2740 (2024)

    2024

  14. [22]

    Tang, Y.B., Yan, K., Tang, Y.X., Liu, J., Xiao, J., Summers, R.M.: Uldor: a uni- versal lesion detector for ct scans with pseudo masks and hard negative example 10 Di et al. mining. In: 2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019). pp. 833–836. IEEE (2019)

    2019

  15. [23]

    Ting, D.S.W., Tan, T.E., Lim, C.T.: Development and validation of a deep learning systemfordetectionofactivepulmonarytuberculosisonchestradiographs:Clinical and technical considerations (2019)

    2019

  16. [24]

    Frontiers in human neuroscience13, 213 (2019)

    Waite, S., Grigorian, A., Alexander, R.G., Macknik, S.L., Carrasco, M., Heeger, D.J., Martinez-Conde, S.: Analysis of perceptual expertise in radiology–current knowledge and a new perspective. Frontiers in human neuroscience13, 213 (2019)

    2019

  17. [25]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Yan, K., Bagheri, M., Summers, R.M.: 3d context enhanced region-based convolu- tional neural network for end-to-end lesion detection. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 511–519. Springer (2018)

    2018

  18. [26]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Yan, K., Tang, Y., Peng, Y., Sandfort, V., Bagheri, M., Lu, Z., Summers, R.M.: Mulan: Multitask universal lesion analysis network for joint lesion detection, tag- ging, and segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention...

    2019

  19. [27]

    Journal of medical imaging5(3), 036501 (2018)

    Yan, K., Wang, X., Lu, L., Summers, R.M.: Deeplesion: automated mining of large- scale lesion annotations and universal lesion detection with deep learning. Journal of medical imaging5(3), 036501 (2018)

    2018

  20. [28]

    Nature Communications15(1), 1131 (2024)

    Ying, H., Liu, X., Zhang, M., Ren, Y., Zhen, S., Wang, X., Liu, B., Hu, P., Duan, L., Cai, M., et al.: A multicenter clinical ai system study for detection and diagnosis of focal liver lesions. Nature Communications15(1), 1131 (2024)

    2024

  21. [29]

    In: Annual Conference on Medical Image Understanding and Analysis

    Yu, H., Fan, D., Song, W.: Gpu-net: Lightweight u-net with more diverse features. In: Annual Conference on Medical Image Understanding and Analysis. pp. 223–

  22. [30]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    Zhu, C., He, Y., Savvides, M.: Feature selective anchor-free module for single-shot object detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 840–849 (2019)

    2019

  23. [31]

    Zlocha, M., Dou, Q., Glocker, B.: Improving retinanet for ct lesion detection with densemasksfromweakrecistlabels.In:InternationalConferenceonMedicalImage Computing and Computer-Assisted Intervention. pp. 402–410. Springer (2019)

    2019

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.