Pith. sign in

REVIEW 3 major objections 6 minor 26 references

Contour priors on three orthogonal CT planes raise semi-automated lesion Dice from 0.67 to 0.88 on an external set of 3865 lesions.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-31 20:52 UTC pith:ESVVZ4FV

load-bearing objection Clean, large-n ablation of prior type × orthogonal planes on nnU-Net; ranking is real under their protocol but partly baked in by GT-derived contours, and the clinical-efficiency claim is still untested. the 3 major comments →

arxiv 2607.24210 v1 pith:ESVVZ4FV submitted 2026-07-27 cs.CV

Effect of User-Prompted Priors on Semi-Automated Cancer Lesion Segmentation in Whole-Body Computed Tomography

Isac Stark , Johan \"Ofverstedt , Elin Lundstr\"om , Simon Ekstr\"om , H{\aa}kan Ahlstr\"om , Joel Kullberg This is my paper
classification cs.CV
keywords Medical ImagingSegmentationDeep LearningSegmentation priorswhole-body CTtumour volumeuser promptsRECIST
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Total tumour volume predicts outcomes better than RECIST diameters, but drawing every lesion by hand is too slow for large trials. This paper shows that feeding a standard 3D segmentation network cheap spatial hints from a user—especially lesion contours drawn on axial, coronal and sagittal slices—sharply improves volumetric accuracy in whole-body CT. Performance rose steadily as the prior went from none, to a point click, to boxes, to contours, and as more orthogonal planes were added. On a large external melanoma test set the best three-plane contour prior reached mean Dice 0.882 versus 0.671 with no prior. The result matters because it offers a practical route to high-quality volumetric ground truth without full manual delineation of every metastasis.

Core claim

More complex spatial user priors consistently improve semi-automated cancer lesion segmentation in whole-body CT. Contour priors supplied on three orthogonal planes achieve the highest performance, reaching mean Dice 0.882 on an external test of 3865 lesions compared with 0.671 for a cropped baseline given no spatial prior. Adding each further orthogonal plane produces statistically significant gains for both bounding-box and contour priors.

What carries the argument

User-prompted spatial priors (points, tight bounding boxes, or maximal-slice contours) encoded as an extra binary channel to a 3D U-Net, supplied on one, two or three orthogonal planes inside a fixed crop around each known lesion.

Load-bearing premise

The priors were simulated from perfect ground-truth masks with only small random shifts, so the claimed gains assume real operators will draw boxes and contours of similar quality.

What would settle it

Have radiologists draw real multi-plane contours and boxes on the same lesions, then compare both Dice scores and wall-clock annotation time against the simulated-prior results and against full manual 3D delineation.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Multi-plane contour prompts can support efficient creation of volumetric ground-truth datasets for trials and radiomics.
  • Richer priors trade operator effort for higher accuracy; simpler priors may still suffice for well-defined lesions.
  • Cascade pipelines that first detect lesions then segment them with orthogonal prompts become more practical.
  • Total tumour volume measurement in oncology studies becomes more feasible without full slice-by-slice manual contouring.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Real interactive timing studies will likely favour two-plane contours over three once operator fatigue and interface friction are counted.
  • The large jump from box to contour implies that boundary shape, not merely extent, is the missing cue for low-contrast metastases.
  • Because the work assumes lesions are already localised, the efficiency claim for total tumour volume still depends on a separate high-recall detector.
  • Orthogonal-plane prompting should generalise to other anisotropic 3D modalities where a single plane under-specifies irregular objects.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper studies how the type and richness of simulated user prompts affect semi-automated 3D lesion segmentation in whole-body CT. Using nnU-Net as a fixed backbone, the authors compare no-prior (crop only), centroid point, bounding-box, and contour priors encoded as extra input channels, with box/contour priors drawn from one, two, or three orthogonal planes. Training uses 3-fold cross-validation on 906 lesions (DeepLesion3D + FDG-PET-CT-Lesions); external evaluation uses the Longitudinal CT cohort (3865 lesions, 300 melanoma patients). The main findings: richer priors monotonically improve Dice (contour-3-plane 0.882 vs crop-only 0.671 on the external set); adding orthogonal planes gives statistically significant gains (Wilcoxon, Bonferroni-corrected); and a perturbation study shows performance degrades gracefully under small random translations of the prior, with contour priors degrading more than boxes. The authors conclude multi-plane contour prompts are the most effective prior for generating volumetric ground truth efficiently.

Significance. If the results hold, this is a useful, well-controlled benchmark for a practically important problem: volumetric lesion segmentation is the bottleneck for TTV-based endpoints that correlate better with survival than RECIST. Particular strengths worth crediting: (i) the nnU-Net backbone and crop are held fixed across all conditions, so differences are attributable to the prior channel alone; (ii) the evaluation uses a large external holdout (n=3865 lesions from 300 patients) disjoint from training; (iii) the authors include a perturbation study of prior misplacement (Table 2) rather than reporting only oracle-quality prompts; (iv) significance is assessed with paired tests and multiple-comparison correction. All datasets are public, so the study is in principle reproducible. The findings quantify a practically relevant trade-off between prompt complexity and accuracy, and the observation that contour priors degrade faster than boxes under misplacement is itself informative for workflow design.

major comments (3)
  1. [§3.2, Table 2, §5] §3.2 / Table 2 / §5 (Conclusions): The headline recommendation — that 3-plane contour priors are the most effective prompt — rests on a prior-error model that does not match how real operator priors fail. The perturbation study translates a perfect ground-truth contour by ±1–3 voxels in-plane and ±1 voxel through-plane. Real operator contours deviate from GT boundaries systematically (over/under-segmentation, irregular boundary noise; inter-observer contour Dice is typically 0.7–0.85), not by rigid sub-voxel shifts of an exact boundary. Table 2 itself shows contour priors degrade much faster than bboxes under these shifts (0.882→0.790 vs 0.791→0.763 at the largest perturbation), i.e., the contour advantage is concentrated in precisely the regime where the error model is least realistic. To be fair, the ordering does not invert even at the largest tested shift (perturbed contour 0.790 ≈ u
  2. [§3.2, Fig. 2, Abstract] §3.2 / Fig. 2 / Abstract: The 'No Prior' condition is not prior-free. All lesions, including the baseline, are cropped to a 6.7 cm cube centered on the GT lesion center (99th-percentile crop). A GT-centered tight crop is itself a strong localization prior — arguably stronger than a loose user-drawn bounding box. The abstract's claim of '0.882 vs 0.671 for the baseline model with no spatial prior' therefore overstates the contrast: the comparison is 'crop-only vs crop+prompt', not 'no prior vs prior'. This matters because the magnitude of the gain (0.21 Dice) is the paper's headline number. Please relabel the baseline (e.g., 'crop-only') in Fig. 2, Table 2, the abstract, and §5, and state explicitly that absolute gains would be smaller against a less localized reference.
  3. [§3.4, Fig. 2] §3.4 / Fig. 2: The Wilcoxon tests treat n=906 and n=3865 lesions as independent paired samples, but the test set contains 3865 lesions from only 300 patients, and the training cohort 906 lesions from ~780 patients. Lesions within a patient (shared scanner, protocol, disease biology, annotator) are correlated, so the effective sample size is smaller and the reported p-values/CIs are anti-conservative. Given the very large effect sizes (0.05–0.2 Dice between conditions) the qualitative conclusions almost certainly stand, but this should be verified or acknowledged — e.g., cluster-robust variance estimation, a patient-level bootstrap, or a mixed-effects model with patient as a random effect. Relatedly, §3.4 states the Bonferroni correction accounts for 'the number of paired samples' (n=906 / n=3865); a Bonferroni factor should be the number of comparisons, not the number of samples. If α wa
minor comments (6)
  1. [§3.1, §5] The 'external' test set is external in cohort but not in cancer type: FDG-PET-CT-Lesions contributes melanoma cases to training and Longitudinal CT is entirely melanoma. §5 acknowledges this obliquely ('both contain lesions from patients suffering from melanoma') but the generalizability framing in the abstract and §5 should be tempered accordingly.
  2. [§4, §5] The test set outperforms cross-validation for every prior (e.g., 0.882 vs 0.830). The authors' explanation (coarse PET-derived GT in part of the training data penalizes CV scores) is plausible, but it would strengthen the paper to verify it directly — e.g., report CV scores split by dataset of origin (DeepLesion3D vs FDG-PET-CT-Lesions).
  3. [§3.3, Eq. (1)] Point-prior evaluation uses a GT centroid click, while training samples points uniformly within the mask (Eq. 1). The centroid is the most informative single point; a brief robustness check with off-center random points at inference would show how sensitive the point-prior condition is to click placement, which is the realistic operator variable.
  4. [§2, §4] No comparison to contemporary promptable baselines (nnInteractive, MedSAM/SegVol, ONCOPILOT) is provided, even though these are cited in §2 and ONCOPILOT reports numbers on related data. A single baseline run of nnInteractive on the test set would substantially strengthen the claim that the proposed prior-encoding scheme is competitive; at minimum, the absence of such a comparison should be stated as a limitation.
  5. [Fig. 2] Fig. 2 error bars are described as 95% CIs of the per-lesion Dice distribution; given n=3865 these are vanishingly small and add little. Patient-level or fold-level variability would be more informative. Also consider reporting median Dice and a small-lesion subgroup, since mean Dice can be driven by large, easy lesions.
  6. [§2, §3.2, §5, Table 1] Notation/typos: 'an mean Dice' (§2); 'but that the optimal prior may, in practice, depend...' has a stray comma (§5); Table 1's voxel-size row is hard to parse (PET and CT resolutions run together). The crop side length of 6.7 cm is given without the corresponding voxel dimensions of the resampled crop, which would help reproducibility.

Circularity Check

1 steps flagged

Standard GT-simulated interactive priors; Dice ranking is empirically measured, not algebraically forced by a fit or self-definition.

specific steps
  1. other [§3.2 User-Prompted Priors; also Table 2 perturbation]
    "Both the bounding box and contour priors were generated from the ground-truth lesion masks by identifying the slice with the largest lesion cross-section. ... For the bounding boxes, the smallest possible fitting rectangle over the original lesion mask was extracted. For the contour, the slice with the maximal area of the lesion mask itself was extracted along the specified axis."

    The spatial priors used as model inputs are direct extractions (tight box or exact maximal-area mask slice) from the same GT masks against which Dice is scored. Richer priors therefore inject strictly more of the evaluation target as an input channel, so the monotonic ranking contour > bbox > point > none is strongly favored by construction of the independent variable. This is ordinary interactive-seg simulation, not a fitted-constant ‘prediction,’ and residual error plus perturbation drops show it is not fully tautological—but it is the only mild circularity-adjacent step.

full rationale

This is an empirical nnU-Net comparison of prompt types on held-out public CT lesion data. The central numbers (mean Dice 0.882 for 3-plane contours vs 0.671 for cropped no-prior on n=3865 external lesions; Wilcoxon gains from adding planes) are measured outputs of trained models, not rearrangements of fitted constants or self-cited uniqueness theorems. Priors are explicitly simulated from ground-truth masks (largest-slice contour / tight box, plus small translations), which is disclosed standard practice in promptable segmentation and creates expected information-content ordering, but does not make Dice equal the prior by construction: the network must still propagate 2D cues into full 3D volumes, and imperfect Dice (well below 1.0) plus larger contour degradation under perturbation show the result is not tautological. No self-citation chain, ansatz smuggling, or renamed known law carries the claim. Mild residual concern only: the independent variable is carved from the same GT used for scoring, so the experiment measures ‘how much GT slice information helps’ more than fully realistic operator error; that is a disclosed limitation, not circular derivation. Score 2.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

Central claim rests on standard supervised segmentation practice plus several domain modeling choices: lesion location already known (crop-first pipeline), GT-derived priors as stand-ins for users, nnU-Net as adequate backbone, and Dice as the performance criterion. No new physical entities; free parameters are engineering choices (crop percentile, translation ranges, Poisson point rate) that affect absolute scores but not the qualitative ranking if held fixed across arms.

free parameters (3)
  • crop_side_length_percentile = 99th percentile ≈ 6.7 cm
    Crop cube set to 99th percentile of training lesion size (~6.7 cm side); larger lesions use sliding window. Affects context and failure modes.
  • point_prior_Poisson_lambda = λ=0.5, offset 1
    Number of click points N ~ Pois(0.5)+1 during training; centroid used at validation. Controls sparsity of point prior.
  • prior_translation_ranges = train ±3x/y, ±1z
    Random prior shifts ±3 in-plane voxels, ±1 through-plane during training; separate test-time perturbation table uses several ranges. Hand-chosen to simulate user error given anisotropic spacing.
axioms (5)
  • domain assumption Lesion center/location is known a priori so fixed cropping around the lesion is valid and surrounding far anatomy can be discarded.
    Stated in Lesion Cropping (§3.2); pipeline does not address detection.
  • ad hoc to paper Ground-truth-derived largest-slice contours/boxes plus small random translations adequately represent user-prompted priors for ranking methods.
    User-Prompted Priors and custom trainers (§3.2–3.3); acknowledged as limitation in Discussion.
  • domain assumption Holding the nnU-Net self-configuring backbone fixed isolates the effect of prior type and plane count.
    Network architecture §3.3; authors note architecture dependence is out of scope.
  • domain assumption Dice coefficient on external manual radiologist contours is a sufficient primary metric for semi-automated ground-truth utility.
    Statistical Evaluation §3.4; no surface-distance or volume-error primary endpoint.
  • domain assumption Manually PET-aligned, slice-interpolated FDG-PET-CT lesion labels are acceptable CT training targets despite coarser native PET spacing.
    Dataset §3.1; authors later attribute train/test gap partly to this.

pith-pipeline@v1.2.0-grok45-kimik3 · 15080 in / 3050 out tokens · 70090 ms · 2026-07-31T20:52:24.619404+00:00 · methodology

0 comments
read the original abstract

In clinical oncology studies, metastatic cancer is commonly evaluated using "Response Evaluation Criteria in Solid Tumors" (RECIST), in which the diameter of up to five lesions is measured and followed over the course of treatment. However, RECIST shows limited correlation with overall survival. Total tumour volume (TTV) is a stronger predictor but typically relies on manual ground-truth segmentation of all lesions, which is time-consuming and requires expert domain knowledge. Semi-automated approaches leveraging user-prompted priors, such as bounding boxes and single-slice contours, as inputs to automated segmentation methods can facilitate the generation of ground-truth segmentations. This work investigates the impact of different user-prompted priors on semi-automated cancer lesion segmentation performance in whole-body computed tomography. Across 3-fold cross-validation and external testing, more complex spatial priors consistently improved performance, with contour priors from three orthogonal planes (axial, coronal and sagittal) achieving the best results. On the external test (n=3865 lesions), this approach achieved a mean Dice score of 0.882, compared to a mean Dice score of 0.671 for the baseline model with no spatial prior. These findings suggest that the use of multi-plane orthogonal user-prompted priors can improve semi-automated tumour lesion segmentation and support efficient generation of high-quality volumetric ground-truth data.

Figures

Figures reproduced from arXiv: 2607.24210 by Elin Lundstr\"om, H{\aa}kan Ahlstr\"om, Isac Stark, Joel Kullberg, Johan \"Ofverstedt, Simon Ekstr\"om.

Figure 1
Figure 1. Figure 1: Schematic overview of the proposed method. Lesions are first extracted [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Mean Dice score across (a) 3-fold cross-validation ( [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

26 extracted references · 9 canonical work pages

  1. [1]

    World Journal of Clinical Oncology 11(2), 53–73 (Feb 2020)

    Aykan, N.F., Özatlı, T.: Objective response rate assessment in oncology: Current situation and future expectations. World Journal of Clinical Oncology 11(2), 53–73 (Feb 2020). https://doi.org/10.5306/wjco.v11.i2.53

    work page doi:10.5306/wjco.v11.i2.53 2020

  2. [2]

    CA: A Cancer Journal for Clinicians 74(3), 229–263 (2024)

    Bray, F., Laversanne, M., Sung, H., Ferlay, J., Siegel, R.L., Soerjomataram, I., Jemal, A.: Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA: A Cancer Journal for Clinicians 74(3), 229–263 (2024). https://doi.org/10.3322/caac.21834 10 I. Stark et al

    doi:10.3322/caac.21834 2022

  3. [3]

    Medical Image Analysis 102, 103525 (May 2025)

    de Grauw, M.J.J., Scholten, E.Th., Smit, E.J., Rutten, M.J.C.M., Prokop, M., van Ginneken, B., Hering, A.: The ULS23 challenge: A baseline model and benchmark dataset for 3D universal lesion segmentation in computed tomography. Medical Image Analysis 102, 103525 (May 2025). https://doi.org/10.1016/j.media.2025. 103525

    doi:10.1016/j.media.2025 2025

  4. [4]

    American Journal of Cancer Research 11(4), 1121–1131 (Apr 2021)

    Delgado, A., Guddati, A.K.: Clinical endpoints in oncology - a primer. American Journal of Cancer Research 11(4), 1121–1131 (Apr 2021)

    2021

  5. [5]

    https://doi.org/10.48550/arXiv.2311

    Du, Y., Bai, F., Huang, T., Zhao, B.: SegVol: Universal and Interactive Volumet- ric Medical Image Segmentation (Feb 2025). https://doi.org/10.48550/arXiv.2311. 13385

    doi:10.48550/arxiv.2311 2025

  6. [6]

    European Journal of Cancer 45(2), 228–247 (Jan 2009)

    Eisenhauer, E.A., Therasse, P., Bogaerts, J., Schwartz, L.H., Sargent, D., Ford, R., Dancey, J., Arbuck, S., Gwyther, S., Mooney, M., Rubinstein, L., Shankar, L., Dodd, L., Kaplan, R., Lacombe, D., Verweij, J.: New response evaluation criteria in solid tumours: Revised RECIST guideline (version 1.1). European Journal of Cancer 45(2), 228–247 (Jan 2009). h...

    doi:10.1016/j.ejca.2008.10.026 2009

  7. [7]

    Cancer Imaging 13(1), 140–149 (Mar 2013)

    Ganeshan, B., Miles, K.A.: Quantifying tumour heterogeneity with CT. Cancer Imaging 13(1), 140–149 (Mar 2013). https://doi.org/10.1102/1470-7330.2013.0015

    arXiv 2013

  8. [8]

    Nature Machine Intelligence 6(11), 1396–1405 (Nov 2024)

    Gatidis, S., Früh, M., Fabritius, M.P., Gu, S., Nikolaou, K., Fougère, C.L., Ye, J., He, J., Peng, Y., Bi, L., Ma, J., Wang, B., Zhang, J., Huang, Y., Heiliger, L., Marinov, Z., Stiefelhagen, R., Egger, J., Kleesiek, J., Sibille, L., Xiang, L., Bendazzoli, S., Astaraki, M., Ingrisch, M., Cyran, C.C., Küstner, T.: Results from the autoPET challenge on full...

    2024

  9. [9]

    Journal of Medical Imaging 5(1), 011005 (Jan 2018)

    Huang, Q., Lu, L., Dercle, L., Lichtenstein, P., Li, Y., Yin, Q., Zong, M., Schwartz, L., Zhao, B.: Interobserver variability in tumor contouring affects the use of ra- diomics to predict mutational status. Journal of Medical Imaging 5(1), 011005 (Jan 2018). https://doi.org/10.1117/1.JMI.5.1.011005

    work page doi:10.1117/1.jmi.5.1.011005 2018

  10. [10]

    Communications Medicine 3(1), 76 (May 2023)

    Hudock, N.L., Mani, K., Khunsriraksakul, C., Walter, V., Nekhlyudov, L., Wang, M., Lehrer, E.J., Hudock, M.R., Liu, D.J., Spratt, D.E., Zaorsky, N.G.: Future trends in incidence and long-term survival of metastatic cancer in the United States. Communications Medicine 3(1), 76 (May 2023). https://doi.org/10.1038/ s43856-023-00304-x

    2023

  11. [11]

    Nature Methods 18(2), 203–211 (Feb 2021)

    Isensee, F., Jaeger, P.F., Kohl, S.A.A., Petersen, J., Maier-Hein, K.H.: nnU- Net: A self-configuring method for deep learning-based biomedical image seg- mentation. Nature Methods 18(2), 203–211 (Feb 2021). https://doi.org/10.1038/ s41592-020-01008-z

    2021

  12. [12]

    https://doi.org/10.48550/arXiv.2503.08373

    Isensee, F., Rokuss, M., Krämer, L., Dinkelacker, S., Ravindran, A., Stritzke, F., Hamm, B., Wald, T., Langenberg, M., Ulrich, C., Deissler, J., Floca, R., Maier- Hein, K.: nnInteractive: Redefining 3D Promptable Segmentation (Mar 2025). https://doi.org/10.48550/arXiv.2503.08373

    Pith/arXiv arXiv doi:10.48550/arxiv.2503.08373 2025

  13. [13]

    https://doi.org/10.48550/arXiv.2304.02643

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollár, P., Girshick, R.: Segment Anything (Apr 2023). https://doi.org/10.48550/arXiv.2304.02643

    Pith/arXiv arXiv doi:10.48550/arxiv.2304.02643 2023

  14. [14]

    Clinical cancer research : an official journal of the American Association for Cancer Research 19(10), 2607–2612 (May 2013)

    Korn, R.L., Crowley, J.J.: Overview: Progression-Free Survival as an Endpoint in Clinical Trials with Solid Tumors. Clinical cancer research : an official journal of the American Association for Cancer Research 19(10), 2607–2612 (May 2013). https://doi.org/10.1158/1078-0432.CCR-12-2934

    work page doi:10.1158/1078-0432.ccr-12-2934 2013

  15. [15]

    https://doi.org/10.57754/FDAT.qwsry-7t837 User-Prompted Priors for Semi-Automated Cancer Lesion Segmentation 11

    Küstner, T., Peisen, F., Gatidis, S., Wagner, A., Megne, O., Othman, A., Sanner, A., Loßau, T., Moltz, J.H., Kohlbrandt, T., Hering, A.: Longitudinal-CT (Sep 2025). https://doi.org/10.57754/FDAT.qwsry-7t837 User-Prompted Priors for Semi-Automated Cancer Lesion Segmentation 11

    doi:10.57754/fdat.qwsry-7t837 2025

  16. [16]

    Nature Communications 15(1), 654 (Jan 2024)

    Ma, J., He, Y., Li, F., Han, L., You, C., Wang, B.: Segment anything in medical images. Nature Communications 15(1), 654 (Jan 2024). https://doi.org/10.1038/ s41467-024-44824-z

    2024

  17. [17]

    npj Precision Oncology 9(1), 121 (Apr 2025)

    Machado, L., Alberge, L., Philippe, H., Ferreres, E., Khlaut, J., Dupuis, J., Le Floch, K., Habip Gatenyo, D., Roux, P., Grégory, J., Ronot, M., Dancette, C., Boeken, T., Tordjman, D., Manceron, P., Hérent, P.: A promptable CT foun- dation model for solid tumor evaluation. npj Precision Oncology 9(1), 121 (Apr 2025). https://doi.org/10.1038/s41698-025-00903-y

    work page doi:10.1038/s41698-025-00903-y 2025

  18. [18]

    Clinical and Translational Oncol- ogy 27(11), 4091–4102 (Nov 2025)

    Mariotti, F., Agostini, A., Borgheresi, A., Marchegiani, M., Zannotti, A., Gia- comelli, G., Pierpaoli, L., Tola, E., Galiffa, E., Giovagnoni, A.: Insights into ra- diomics: A comprehensive review for beginners. Clinical and Translational Oncol- ogy 27(11), 4091–4102 (Nov 2025). https://doi.org/10.1007/s12094-025-03939-5

    work page doi:10.1007/s12094-025-03939-5 2025

  19. [19]

    Journal of Clinical Oncology 41(15), 2706–2712 (May 2023)

    Merino, M., Kasamon, Y., Theoret, M., Pazdur, R., Kluetz, P., Gormley, N.: Ir- reconcilable Differences: The Divorce Between Response Rates, Progression-Free Survival, and Overall Survival. Journal of Clinical Oncology 41(15), 2706–2712 (May 2023). https://doi.org/10.1200/JCO.23.00225

    work page doi:10.1200/jco.23.00225 2023

  20. [20]

    eClinicalMedicine 76 (Oct 2024)

    Mittal, A., Kim, M.S., Dunn, S., Wright, K., Gyawali, B.: Frequently asked questions on surrogate endpoints in oncology-opportunities, pitfalls, and the way forward. eClinicalMedicine 76 (Oct 2024). https://doi.org/10.1016/j.eclinm.2024. 102824

    work page doi:10.1016/j.eclinm.2024 2024

  21. [21]

    Seminars in Radiation Oncology 31(3), 186–199 (Jul 2021)

    Pirasteh, A., Lovrec, P., Pedrosa, I.: Imaging and its Impact on Defining the Oligometastatic State. Seminars in Radiation Oncology 31(3), 186–199 (Jul 2021). https://doi.org/10.1016/j.semradonc.2021.03.006

    work page doi:10.1016/j.semradonc.2021.03.006 2021

  22. [22]

    Annals of Oncology 27(3), 373–378 (Mar 2016)

    Saad, E.D., Buyse, M.: Statistical controversies in clinical research: End points other than overall survival are vital for regulatory approval of anticancer agents. Annals of Oncology 27(3), 373–378 (Mar 2016). https://doi.org/10.1093/annonc/ mdv562

    doi:10.1093/annonc/ 2016

  23. [23]

    BMC Medical Imaging 24(1), 241 (Sep 2024)

    Sun, J., Chen, K., He, Z., Ren, S., He, X., Liu, X., Peng, C.: Medical image analysis using improved SAM-Med2D: Segmentation and classification perspec- tives. BMC Medical Imaging 24(1), 241 (Sep 2024). https://doi.org/10.1186/ s12880-024-01401-6

    2024

  24. [24]

    In: Del Bue, A., Canton, C., Pont-Tuset, J., Tommasi, T

    Wang, H., Guo, S., Ye, J., Deng, Z., Cheng, J., Li, T., Chen, J., Su, Y., Huang, Z., Shen, Y., Fu, B., Zhang, S., He, J., Qiao, Y.: SAM-Med3D: Towards General- Purpose Segmentation Models for Volumetric Medical Images. In: Del Bue, A., Canton, C., Pont-Tuset, J., Tommasi, T. (eds.) Computer Vision – ECCV 2024 Workshops. pp. 51–67. Springer Nature Switzerl...

    work page doi:10.1007/978-3-031-91721-9_4 2024

  25. [25]

    Journal of Medical Imaging 5(3), 036501 (Jul 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 Imaging 5(3), 036501 (Jul 2018). https://doi.org/10.1117/1. JMI.5.3.036501

    doi:10.1117/1 2018

  26. [26]

    Stark et al

    Zeeuw, J.M., Kemna, R., Ali, M., van Eck, S., Wesdorp, N.J., Roor, J., van Waes- berghe, J.H.T.M., van den Bergh, J.E., Nota, I.M.G.C., Moos, S.I., van Dieren, S., van Amerongen, M.J., Bond, M.J.G., Chapelle, T., van Dam, R.M., Engel- brecht, M.R.W., Gerhards, M.F., Grunhagen, D.J., van Gulik, T.M., Hermans, J.J., de Jong, K.P., Klaase, J.M., Kok, N.F.M.,...

    arXiv 2025