Pith. sign in

REVIEW 3 major objections 5 minor 49 references

Lesion Detection in CT with Frozen Self-Distilled Features: SALT, a Spatially Adaptive Label-Guided Temperature

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Sharpening the teacher temperature only inside lesion patches makes frozen CT features far more lesion-concentrated and improves 3D detection.

desk verdict A genuinely new conditioning mechanism for self-distillation, tested with an honest but single-seed controlled comparison; the mechanism is plausible and the paper is transparent enough to merit refereeing. read the letter →

arxiv 2608.05100 v1 pith:MX4AP2PG submitted 2026-08-05 cs.CV

classification cs.CV
keywords self-supervisedlearningself-distillationweaksupervisionlesiondetectioncomputedtomographyfrozenrepresentationsteachertemperaturespatiallyadaptiveconditioning
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 claims that spatially uniform self-supervised pretraining under-weights small structures like CT lesions, because a lesion a few patches wide contributes no more to the loss than the surrounding tissue. Its method, SALT, moves the guidance from the view to the target: during pretraining only, weak box-derived labels define a compact patch region inside which the teacher's softmax temperature is sharpened and the masked-patch loss is up-weighted. Against a ViT-B/14 baseline that is identical except for this target-level conditioning, SALT raises the lesion/background separation margin from 0.105 to 0.449 and pooled 3D catch-once detection from 0.394 to 0.508 CPM, with the gain concentrated on sub-6mm lesions. The point of the claim is that labels are needed once at pretraining time, and the frozen encoder then serves as a plain feature extractor for lightweight detection.

What carries the argument

The load-bearing object is the spatial indicator $w(p) = \mathbb{1}[p \in \mathrm{region}]$ on the encoder's patch grid, which turns a single scalar teacher temperature into a per-token temperature field. Box-derived masks are pooled onto the 16×16 patch grid, the connected group of lesion patches nearest the view center is kept and dilated into a compact square, and inside that region the teacher's softmax temperature is sharpened and the patch loss is up-weighted; everything else in the DINOv2/iBOT recipe, including masking, centering, and EMA momentum, is unchanged. The controlled comparison isolates this mechanism: the SALT ViT-B and DINO-LG ViT-B differ only in these target-level operations, sharing architecture, pretraining data, schedule, seed, and lesion-guided cropping. The detector that exposes the effect reads frozen patch tokens from four depths and trains only a multi-depth CenterNet-style head, so any difference is attributable to the frozen features.

What would settle it

Pretrain the matched ViT-B pair (SALT vs DINO-LG) with five seeds each on the same eight-cohort store and recompute the separation margin and pooled CPM; if the mean SALT-minus-DINO-LG difference collapses toward zero or changes sign, the central claim fails. Also, sweep the lesion temperature $\tau_\ell$ with the loss weight fixed at 1.0: if the separation margin does not rise with sharpening, the attributed mechanism is wrong. A simpler check is to place the conditioning region at a random location rather than the lesion; if CPM still rises, the effect is not label-driven.

Watch

Extended reading notes

Core claim

The central discovery is that target-level conditioning, not view-level guidance, is what makes a frozen representation lesion-centered. In self-distillation, the teacher temperature and per-patch loss weight are normally uniform; SALT replaces the temperature with $\tau(p) = \tau_{\mathrm{base}} - (\tau_{\mathrm{base}} - \tau_\ell)\,\mathbb{1}[p \in \mathrm{region}]$, applying a sharp temperature $\tau_\ell = 0.025$ inside a compact, dilated region derived from box masks, and multiplies the masked-patch loss by three there. The teacher still encodes the intact view, and the centering statistics are accumulated before the sharpening, so the targets remain calibrated. Evaluated with a frozen multi-depth ViT descriptor and a lightweight CenterNet-style head, the conditioned ViT-B reaches a separation margin of 0.449 and CPM 0.508, versus 0.105 and 0.394 for the matched DINO-LG baseline without target conditioning; the improvement is large on sub-6mm lesions and on two of four cohorts.

Load-bearing premise

The controlled comparison rests on a single pretraining seed per configuration and representation-level geometry measured on one random volume per cohort, so the headline gap between SALT and DINO-LG has no sampling error bars and could be a favorable seed.

Editorial extensions

If this is right

  • Small-lesion detection in CT can be improved without fine-tuning the encoder or using labels at inference; the conditioning cost is paid once at pretraining.
  • The +0.136 CPM gain on sub-6mm lesions is exactly where the mechanism predicts it: where a lesion spans two or three patches, the sharpened teacher target supplies the only local pressure to encode it distinctly.
  • The benefit is not uniform across cohorts: conditioning raises CPM by +0.144 on UniToChest and +0.210 on NIH-Lymph and is slightly negative on NLST-seg and HCUCH, so the pooled +0.11 should be read as cohort-dependent.
  • The representation-level effect precedes detection: the separation margin changes by roughly a factor of four before any detector is attached, while linear-probe AUROC is near-ceiling for all backbones, so decodability is not the same as concentration.
  • Because the conditioning is expressed through a spatial indicator rather than label semantics, the formulation admits any weak spatial annotation, not just lesions.

Reading between the lines

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

  • Extension: a direct test of the mechanism on other small, low-contrast targets whose footprint is a few patches, such as polyps, microbleeds, or fractures, would show whether the conditioning transfers beyond lesions.
  • Extension: sweeping $\tau_\ell$ and the patch-loss weight separately would resolve which part of the target conditioning carries the effect; the paper leaves them unseparated.
  • Extension: running the matched ViT-B pair with multiple pretraining seeds and reporting the distribution of the separation margin and CPM gap would turn the headline numbers from point estimates into an effect size with variance.
  • Extension: the detector-free re-identification probe hints that the frozen conditional features may support longitudinal lesion matching without registration; with only 28 paired lesions the paper does not claim this, but a larger paired cohort could test it.
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

3 major / 5 minor

Summary. The paper introduces SALT, a self-distillation pretraining objective that uses weak box-derived lesion labels to define a compact spatial region on the patch grid where the teacher softmax temperature is sharpened and the masked-patch loss is up-weighted. The authors freeze the pretrained encoder and train a lightweight CenterNet-style detection head on four CT cohorts, comparing SALT against a matched DINO-LG baseline (same architecture, data, schedule, seed, and lesion-guided cropping, but no target conditioning) and against external frozen backbones. They report that SALT improves lesion/background patch separation margin (0.449 vs 0.105) and pooled 3D detection CPM (0.508 vs 0.394), with the detection gain concentrated on sub-6mm lesions. The paper also reports a longitudinal re-identification probe and is unusually transparent about its limitations: one pretraining seed, test-set operating-point selection, one volume per cohort for geometry, and negative per-cohort differences on two of four cohorts.

Significance. If the central claim holds, SALT is a simple and potentially general mechanism for injecting weak spatial labels into self-supervised pretraining at the target level, and the paper provides a clean matched-pair design that isolates target conditioning from view-level guidance. The paper's strengths are its transparency (negative cohorts reported, best-case F1 explicitly labeled, threshold-free CPM as primary metric, one-seed limitation stated) and its controlled baseline, re-run under identical conditions. The proposed mechanism is novel in combining per-token temperature modulation with an external spatial label inside self-distillation. However, the causal claim currently rests on a single seed and on a patch-geometry measurement from four volumes, so the significance as established is moderate until variance is quantified.

major comments (3)
  1. [§5.5, §6] The paper's central claim—that target conditioning causes the 0.449 vs 0.105 separation-margin gap and the +0.11 CPM gain—rests entirely on a matched pair of pretrained encoders, each trained with one seed. The manuscript itself states in Section 5.5: 'With one pretraining seed and no interval on the per-cohort differences we cannot say whether the two negative signs reflect noise, a ceiling, or a real cost of conditioning on those distributions.' This is a load-bearing limitation: self-distillation pretraining is stochastic, and without multiple seeds or a variance estimate, the observed gaps could be due to run-to-run variation. I request that the authors provide results from at least three (ideally five) pretraining seeds for the SALT ViT-B and DINO-LG ViT-B pair, reporting the range or confidence intervals on sep and CPM.
  2. [§5.3, Table 4] The patch-geometry evidence (sep and bg-sim) is computed from one randomly drawn volume per cohort, the same four volumes for every backbone. With n=4 volumes, the fourfold gap in separation margin between SALT ViT-B (0.449) and DINO-LG ViT-B (0.105) has no sampling error bars; a single unusual volume could drive the comparison. Since this representation-level result is the mechanism-level support for the detection gains, the authors should report per-volume values and bootstrap or other confidence intervals, or use more volumes if available.
  3. [§5.5, Table 6] The pooled +0.11 CPM gain in Table 5 is almost entirely a UniToChest effect: UniToChest supplies 3008 of 3983 lesions, and over the other three cohorts the SALT-vs-DINO-LG residual is only +0.021 CPM. The paper acknowledges this in Section 5.5, but the abstract and contribution list state the +0.11 pooled gain without this qualifier. I recommend that the pooled claim be accompanied by a leave-one-cohort-out analysis, or explicitly state that the effect is established primarily on UniToChest and NIH-Lymph, with negative differences on NLST-seg and HCUCH that cannot currently be separated from noise (as in major comment 1).
minor comments (5)
  1. [§3.3, §6] The values τ_l=0.025, dilation=3, and patch-loss weight ×3 are given without a sensitivity analysis, and Section 6 notes that the sharpened temperature and up-weighted loss are not separated. A short ablation or a statement on how these values were chosen would increase confidence that they are not tuned to the test set.
  2. [References, footnote] Reference [34] is an unfinished 'TODO' entry ('Tübingen lung/node/liver ct cohort, TODO'), and the author footnote contains a TODO for email addresses and ORCIDs; these should be completed before publication.
  3. [§5.3] The linear-probe AUROC values (≥0.99 for all backbones) are computed with 5-fold cross-validation on patches; because patches within a slice are spatially correlated, random patch-level folds may inflate AUROC due to information leakage. Consider evaluating the probe with slice- or volume-level splits, or at least acknowledge this.
  4. [§5.4] The operating point that maximizes F1 is chosen on the test set; the manuscript explicitly labels these as best-case, which is transparent, but if space permits, a validation split would allow non-best-case estimates and would strengthen the F1 comparisons.
  5. [§5.7, Table 12] The re-identification probe uses the ground-truth baseline center as the query seed, so the probe is not fully automatic; the text explains this in the protocol, but the phrase 'detection-free' in the table title might be misread as fully unsupervised, and a clarification would help.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the controlled SALT/DINO-LG comparison is implemented in-paper, and no fitted parameter is renamed as a prediction.

full rationale

The paper's derivation chain is self-contained. SALT's conditioning (Eq. 1) changes the teacher softmax temperature and patch-loss weight on a weak-label-derived region during pretraining, and the claimed effects (separation margin, pooled CPM, size-stratified gains) are measured on frozen encoders under a fixed protocol. The controlled DINO-LG baseline is trained in-paper under the stated identical architecture, data, schedule, and seed, so the self-citation [8] is not load-bearing: the comparison does not rely on the cited paper's numbers. No SALT hyperparameter is fitted to downstream detection; the paper explicitly discloses that the F1 operating threshold is selected on the test set, and CPM is threshold-free, so the primary detection metric is not a fitted prediction. The linear probe AUROC >= 0.99 for every backbone shows a lesion direction exists everywhere, so the separation-margin gap is a distinct cosine-geometry measure rather than a relabeled training signal. The notable limitations are statistical: one pretraining seed and no interval on per-cohort differences (Sec. 5.5), one random volume per cohort for the geometry table (Sec. 5.3), and no held-out validation split for operating-point selection. These are correctness and robustness concerns, not circularity: no equation reduces to its own input, and no fitted parameter is renamed as a prediction.

Assumptions & free parameters 7 free parameters · 7 assumptions · 0 invented entities

The central claim rests on standard self-distillation machinery (DINOv2/iBOT), on the assumption that weak box-derived masks mark the lesion faithfully, on patient-disjoint splits taken on trust from store metadata, on the representativeness of one random volume per cohort, on a single seed, and on the clinical adequacy of catch-once matching. No new physical entities are introduced. The hand-chosen hyperparameters that tune the conditioning are listed above.

free parameters (7)
  • Lesion teacher temperature tau_l = 0.025
    Hand-chosen sharpening temperature inside the lesion region; directly defines the target conditioning in Eq. (1). No sweep reported.
  • Lesion patch-loss weight multiplier = 3.0
    Hand-chosen up-weight factor for masked patches inside the lesion region; not swept separately from the temperature.
  • Region dilation (patches) = 3
    Hand-chosen dilation of the compact lesion region; affects which patches receive sharpened targets.
  • Teacher temperature base schedule tau_base = 0.04 to 0.07 with 15k warmup
    Standard DINOv2 warmup schedule, but still a chosen hyperparameter of the target distribution.
  • Presence gate threshold = 0.5
    Train-time default for slice-presence gating; downstream operating point, same across backbones.
  • Detection match radius = 1.5 cells on 36-grid
    Hand-chosen tolerance for catch-once localization; affects recall and CPM values.
  • Focal loss alpha/beta and losses lambda/gamma = alpha=2, beta=4, lambda=1, gamma=0.1
    Hand-chosen head loss weights; applied identically to all backbones, so they affect absolute numbers but not the matched comparison.
assumptions (7)
  • domain assumption The DINOv2/iBOT self-distillation objective remains well-behaved when the teacher temperature and patch-loss weight are spatially modulated by a label region.
    Section 3.2 states that the teacher encodes the intact view and that centering statistics are computed before temperature is applied, but no collapse or stability analysis is provided for the modified target.
  • domain assumption Weak box-derived masks, pooled to the patch grid and reduced to a compact region, mark the lesion tissue faithfully enough to benefit pretraining.
    Section 3.2 converts boxes to masks and keeps only the connected group nearest the view center; label noise is not measured.
  • domain assumption The train/test split is patient-disjoint as stated by the store metadata.
    Section 5.1 asserts this from metadata, without an independent verification protocol.
  • ad hoc to paper A single randomly drawn volume per cohort is representative for the patch-separability measurement.
    Section 5.3 draws one volume per cohort and averages over four volumes; no error bars are reported.
  • ad hoc to paper One pretraining seed per configuration is representative for the matched-pair comparison.
    Section 5.5 concedes that one seed cannot separate noise from a real cost on the cohorts with negative gains.
  • domain assumption Catch-once matching with the stated tolerance is an adequate proxy for clinically useful detection.
    Sections 4.4 and 5.4 define the protocol; the paper acknowledges that it does not score volumetric delineation.
  • domain assumption External backbones' pretraining corpora do not overlap the test volumes used here.
    Section 5.2 notes only LiTS overlap with MedDINOv3's CT-3M and does not audit the other public corpora or the private Curia corpus.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Lesion Detection in CT with Frozen Self-Distilled Features: SALT, a Spatially Adaptive Label-Guided Temperature." pith.science (2026). https://pith.science/paper/MX4AP2PG

@misc{pith2026260805100,
  author       = {Pith},
  title        = {Pith review of: Lesion Detection in CT with Frozen Self-Distilled Features: SALT, a Spatially Adaptive Label-Guided Temperature},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MX4AP2PG}},
  note         = {Machine review of arXiv:2608.05100}
}
read the original abstract

Self-supervised pretraining objectives are spatially uniform: the teacher temperature and the per-patch loss weight are identical everywhere in the image, so a lesion a few patches wide contributes no more to the training signal than the surrounding parenchyma. Prior work biases the views toward annotated regions, which changes what the model sees but adds no pressure on the objective. We instead condition the targets of self-distillation, a method we call SALT (Spatially Adaptive Label-guided Temperature). Weak, box-derived labels, available only during pretraining, define a compact region on the encoder's patch grid, inside which the teacher's softmax temperature is sharpened and the masked-patch loss is up-weighted. The objectives, the masking policy and the centering statistics are otherwise unchanged, and at every downstream use the encoder is a plain feature extractor with no labels and no conditioning. We evaluate by freezing the encoder and training only a lightweight multi-depth CenterNet-style head, detecting lesions in 3D on four CT cohorts, and we isolate the mechanism against a backbone identical in architecture, pretraining data, schedule and label-guided cropping but with no target conditioning. We report patch-level separability, 3D detection stratified by cohort and by lesion size, box quality, and a detector-free probe in which a single frozen patch embedding re-identifies a lesion in a follow-up scan without registration, masks or fine-tuning. Because the conditioning is expressed through a spatial indicator rather than through label semantics, the formulation admits any weak spatial annotation; we instantiate and validate it for lesions.

Figures

Figures reproduced from arXiv: 2608.05100 by the authors.

Figure 1
Figure 1. Lesion-conditioned pretraining. The teacher encodes unmasked global views, the student masked globals and [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Frozen-encoder detector. Multi-depth ViT patch tokens ( [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Detector output over five consecutive slices of a lung lesion (NLST-seg). Rows: center-response heatmap; [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Per-patch cosine similarity to the lesion prototype on one UniToChest slice (lesion outline white). Left to [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Detection-free re-identification example (frozen SALT ViT-L; patient [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 40 canonical work pages

  1. [1]

    Objects as points.arXiv preprint arXiv:1904.07850, 2019

    Xingyi Zhou, Dequan Wang, and Philipp Krähenbühl. Objects as points.arXiv preprint arXiv:1904.07850, 2019

  2. [2]

    Reduced lung-cancer mortality with low-dose computed tomo- graphic screening.New England Journal of Medicine, 365(5):395–409, 2011

    National Lung Screening Trial Research Team. Reduced lung-cancer mortality with low-dose computed tomo- graphic screening.New England Journal of Medicine, 365(5):395–409, 2011

  3. [3]

    Eisenhauer, Patrick Therasse, Jan Bogaerts, et al

    Elizabeth A. Eisenhauer, Patrick Therasse, Jan Bogaerts, et al. New response evaluation criteria in solid tumours: Revised RECIST guideline (version 1.1).European Journal of Cancer, 45(2):228–247, 2009

  4. [4]

    Williams and Trafton Drew

    Lauren H. Williams and Trafton Drew. Non-conscious detection of “missed” lung nodules by radiologists: Expanding the boundaries of successful processing during the visual assessment of chest CT scans.Radiology, 2024

  5. [5]

    MedDINOv3: How to adapt vision foundation models for medical image segmentation?arXiv preprint arXiv:2509.02379, 2025

    Yaowei Li et al. MedDINOv3: How to adapt vision foundation models for medical image segmentation?arXiv preprint arXiv:2509.02379, 2025

  6. [6]

    Curia: A multi-modal foundation model for radiology.arXiv preprint arXiv:2509.06830, 2025

    Corentin Dancette et al. Curia: A multi-modal foundation model for radiology.arXiv preprint arXiv:2509.06830, 2025

  7. [7]

    Curia-2: Scaling self-supervised learning for radiology foundation models.arXiv preprint arXiv:2604.01987, 2026

    Adriel Saporta et al. Curia-2: Scaling self-supervised learning for radiology foundation models.arXiv preprint arXiv:2604.01987, 2026

  8. [8]

    Gokmen, Evan W

    Mahmut S. Gokmen, Evan W. Damron, Mitchell A. Klusty, Caroline N. Leach, Emily B. Collier, and V . K. Bum- gardner. DINO-LG: A task-specific DINO model for coronary calcium scoring.arXiv preprint arXiv:2411.07976, 2024

Show all 49 references
  1. [9]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. InInternational Conference on Machine Learning (ICML), pages 1597–1607, 2020

  2. [10]

    Momentum contrast for unsupervised visual representation learning

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9729–9738, 2020

  3. [11]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16000–16009, 2022

  4. [12]

    Gotway, and Jianming Liang

    Zongwei Zhou, Vatsal Sodha, Jiaxuan Pang, Michael B. Gotway, and Jianming Liang. Models genesis.Medical Image Analysis, 67:101840, 2021

  5. [13]

    Emerging properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. InIEEE/CVF International Conference on Computer Vision (ICCV), pages 9650–9660, 2021

  6. [14]

    iBOT: Image BERT pre-training with online tokenizer

    Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. iBOT: Image BERT pre-training with online tokenizer. InInternational Conference on Learning Representations (ICLR), 2022

  7. [15]

    DINOv2: Learning robust visual features without supervision.Transactions on Machine Learning Research (TMLR), 2024

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, et al. DINOv2: Learning robust visual features without supervision.Transactions on Machine Learning Research (TMLR), 2024

  8. [16]

    Robson, et al

    Xueyan Mei, Zelong Liu, Philip M. Robson, et al. RadImageNet: An open radiologic deep learning research dataset for effective transfer learning.Radiology: Artificial Intelligence, 4(5):e210315, 2022

  9. [17]

    Roth, Bennett Landman, Daguang Xu, Vishwesh Nath, and Ali Hatamizadeh

    Yucheng Tang, Dong Yang, Wenqi Li, Holger R. Roth, Bennett Landman, Daguang Xu, Vishwesh Nath, and Ali Hatamizadeh. Self-supervised pre-training of swin transformers for 3D medical image analysis. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2...

  10. [18]

    V o, Maximilian Seitzer, et al

    Oriane Siméoni, Huy V . V o, Maximilian Seitzer, et al. DINOv3.arXiv preprint arXiv:2508.10104, 2025

  11. [19]

    Ke Yan, Xiaosong Wang, Le Lu, and Ronald M. Summers. DeepLesion: Automated mining of large-scale lesion annotations and universal lesion detection with deep learning.Journal of Medical Imaging, 5(3):036501, 2018

  12. [20]

    Ke Yan, Mohammadhadi Bagheri, and Ronald M. Summers. 3D context enhanced region-based convolutional neural network for end-to-end lesion detection. InMedical Image Computing and Computer Assisted Intervention (MICCAI), pages 511–519, 2018

  13. [21]

    Ke Yan, Youbao Tang, Yifan Peng, Veit Sandfort, Mohammadhadi Bagheri, Zhiyong Lu, and Ronald M. Summers. MULAN: Multitask universal lesion analysis network for joint lesion detection, tagging, and segmentation. In Medical Image Computing and Computer Assisted Intervention (MIC...

  14. [22]

    Harrison, Dakai Jin, Jing Xiao, and Le Lu

    Ke Yan, Jinzheng Cai, Adam P. Harrison, Dakai Jin, Jing Xiao, and Le Lu. Learning from multiple datasets with heterogeneous and partial labels for universal lesion detection in CT.IEEE Transactions on Medical Imaging, 40(10):2759–2770, 2021

  15. [23]

    Jäger, Fabian Isensee, and Klaus H

    Michael Baumgartner, Paul F. Jäger, Fabian Isensee, and Klaus H. Maier-Hein. nnDetection: A self-configuring method for medical object detection. InMedical Image Computing and Computer Assisted Intervention (MICCAI), pages 530–539, 2021

  16. [24]

    Arnaud A. A. Setio, Alberto Traverso, Thomas de Bel, Moira S. N. Berens, et al. Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in CT images: The LUNA16 challenge. Medical Image Analysis, 42:1–13, 2017

  17. [25]

    Selvaraju, Karan Desai, Justin Johnson, and Nikhil Naik

    Ramprasaath R. Selvaraju, Karan Desai, Justin Johnson, and Nikhil Naik. CASTing your model: Learning to localize improves self-supervised representations. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11058–11067, 2021

  18. [26]

    Crafting better contrastive views for siamese representation learning

    Xiangyu Peng, Kai Wang, Zheng Zhu, Mang Wang, and Yang You. Crafting better contrastive views for siamese representation learning. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16031–16040, 2022

  19. [27]

    Supervised contrastive learning

    Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. InAdvances in Neural Information Processing Systems (NeurIPS), 2020

  20. [28]

    Curriculum temperature for knowledge distillation

    Zheng Li, Xiang Li, Lingfeng Yang, Borui Zhao, Renjie Song, Lei Luo, Jun Li, and Jian Yang. Curriculum temperature for knowledge distillation. InAAAI Conference on Artificial Intelligence, pages 1504–1512, 2023

  21. [29]

    SAMKD: Spatial-aware adaptive masking knowledge distillation for object detection.arXiv preprint arXiv:2501.07101, 2025

    Duo Wang et al. SAMKD: Spatial-aware adaptive masking knowledge distillation for object detection.arXiv preprint arXiv:2501.07101, 2025

  22. [30]

    UniToChest: A lung image dataset for segmentation of cancerous nodules on CT scans

    Hafiza Ayesha Hoor Chaudhry, Riccardo Renzulli, Daniele Perlo, Francesca Santinelli, Stefano Tibaldi, Carmen Cristiano, Marco Grosso, Giorgio Limerutti, Attilio Fiandrotti, Marco Grangetto, and Paolo Fonio. UniToChest: A lung image dataset for segmentation of cancerous nodules...

  23. [31]

    The liver tumor segmentation benchmark (LiTS).Medical Image Analysis, 84:102680, 2023

    Patrick Bilic, Patrick Christ, Hongwei Bran Li, et al. The liver tumor segmentation benchmark (LiTS).Medical Image Analysis, 84:102680, 2023

  24. [32]

    Roth, Le Lu, Ari Seff, Kevin M

    Holger R. Roth, Le Lu, Ari Seff, Kevin M. Cherry, Joanne Hoffman, Shijun Wang, Jiamin Liu, Evrim Turkbey, and Ronald M. Summers. A new 2.5D representation for lymph node detection using random sets of deep convolutional neural network observations. InMedical Image Computing an...

  25. [33]

    Eyssautier, Felipe Bravo- Bahamóndez, Nicolás Sanhueza, Paulina Gallardo-Badilla, Francisca Caro-Flores, Camila Ormeño-Candia, Felipe Santander, Nicolás Pérez, María M

    Roberto Rojas-Pizarro, Constanza Vásquez-Venegas, Gonzalo Pereira, María F. Eyssautier, Felipe Bravo- Bahamóndez, Nicolás Sanhueza, Paulina Gallardo-Badilla, Francisca Caro-Flores, Camila Ormeño-Candia, Felipe Santander, Nicolás Pérez, María M. Molina, Gonzalo Rojas, Steffen H...

  26. [34]

    Todo: Tübingen lung/node/liver ct cohort, TODO

    TODO. Todo: Tübingen lung/node/liver ct cohort, TODO

  27. [35]

    Leira, and Thomas Langø

    David Bouget, André Pedersen, Johanna Vanel, Haakon O. Leira, and Thomas Langø. Mediastinal lymph nodes segmentation using 3D convolutional neural network ensembles and anatomical priors guiding.Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualizati...

  28. [36]

    Vision transformers need registers

    Timothée Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In International Conference on Learning Representations (ICLR), 2024

  29. [37]

    Maier-Hein, et al

    Nicholas Heller, Fabian Isensee, Klaus H. Maier-Hein, et al. The state of the art in kidney and kidney tumor segmentation in contrast-enhanced CT imaging: Results of the KiTS19 challenge.Medical Image Analysis, 67:101821, 2021

  30. [38]

    AMOS: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation

    Yuanfeng Ji, Haotian Bai, Chongjian Ge, et al. AMOS: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation. InAdvances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track, 2022

  31. [39]

    WORD: A large scale dataset, benchmark and clinical applicable study for abdominal organ segmentation from CT image.Medical Image Analysis, 82:102642, 2022

    Xiangde Luo, Wenjun Liao, Jianghong Xiao, et al. WORD: A large scale dataset, benchmark and clinical applicable study for abdominal organ segmentation from CT image.Medical Image Analysis, 82:102642, 2022

  32. [40]

    Eugenio Igelsias, Martin Styner, Thomas Langerak, and Arno Klein

    Bennett Landman, Zhoubing Xu, J. Eugenio Igelsias, Martin Styner, Thomas Langerak, and Arno Klein. MICCAI multi-atlas labeling beyond the cranial vault: Workshop and challenge. InMICCAI Multi-Atlas Labeling Beyond Cranial Vault Workshop Challenge, 2015. 17 SALT: Spatially Adap...

  33. [41]

    Emre Kavur, N

    A. Emre Kavur, N. Sinem Gezer, Mustafa Barı¸ s, et al. CHAOS challenge: Combined (CT-MR) healthy abdominal organ segmentation.Medical Image Analysis, 69:101950, 2021

  34. [42]

    AbdomenCT-1K: Is abdominal organ segmentation a solved problem?IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6695–6714, 2022

    Jun Ma, Yao Zhang, Song Gu, et al. AbdomenCT-1K: Is abdominal organ segmentation a solved problem?IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6695–6714, 2022

  35. [43]

    The medical segmentation decathlon.Nature Communications, 13:4128, 2022

    Michela Antonelli, Annika Reinke, Spyridon Bakas, et al. The medical segmentation decathlon.Nature Communications, 13:4128, 2022

  36. [44]

    Blaine Rister, Darvin Yi, Kaushik Shivakumar, Tomomi Nobashi, and Daniel L. Rubin. CT-ORG: A new dataset for multiple organ segmentation in computed tomography.Scientific Data, 7:381, 2020

  37. [45]

    Meyer, et al

    Jakob Wasserthal, Hanns-Christian Breit, Manfred T. Meyer, et al. TotalSegmentator: Robust segmentation of 104 anatomic structures in CT images.Radiology: Artificial Intelligence, 5(5):e230024, 2023

  38. [46]

    AbdomenAtlas: A large-scale, detailed-annotated, and multi-center dataset for efficient transfer learning and open algorithmic benchmarking.Medical Image Analysis, 97:103285, 2024

    Wenxuan Li, Chongyu Qu, Xiaoxi Chen, et al. AbdomenAtlas: A large-scale, detailed-annotated, and multi-center dataset for efficient transfer learning and open algorithmic benchmarking.Medical Image Analysis, 97:103285, 2024

  39. [47]

    Roth, Le Lu, Amal Farag, Hoo-Chang Shin, Jiamin Liu, Evrim B

    Holger R. Roth, Le Lu, Amal Farag, Hoo-Chang Shin, Jiamin Liu, Evrim B. Turkbey, and Ronald M. Summers. DeepOrgan: Multi-level deep convolutional networks for automated pancreas segmentation. InMedical Image Computing and Computer Assisted Intervention (MICCAI), pages 556–564, 2015

  40. [48]

    ImageNet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255, 2009

  41. [49]

    Google Landmarks Dataset v2: A large-scale benchmark for instance-level recognition and retrieval

    Tobias Weyand, André Araujo, Bingyi Cao, and Jack Sim. Google Landmarks Dataset v2: A large-scale benchmark for instance-level recognition and retrieval. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2575–2584, 2020. 18

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.