Towards Interactive Lesion Segmentation in Whole-Body PET/CT with Promptable Models

Fabian Isensee; Klaus H. Maier-Hein; Maximilian Rokuss; Yannick Kirchhoff

REVIEW 3 major objections 4 minor 13 references

The paper argues that encoding user clicks as Euclidean distance transforms instead of Gaussian kernels improves interactive lesion segmentation in whole-body PET/CT.

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 ↗

Adding click prompts encoded as Euclidean distance transforms to an nnU-Net improves interactive whole-body PET/CT lesion segmentation over Gaussian encodings and baseline autoPET III models.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A useful ablation of click encodings for interactive PET/CT segmentation, but the headline ensemble result is not actually reported. the 3 major comments →

arxiv 2508.21680 v1 pith:LK4BDFFQ submitted 2025-08-29 cs.CV

Towards Interactive Lesion Segmentation in Whole-Body PET/CT with Promptable Models

Maximilian Rokuss , Yannick Kirchhoff , Fabian Isensee , Klaus H. Maier-Hein This is my paper

classification cs.CV

keywords interactive segmentationlesion segmentationPET/CTpromptable modelsEuclidean Distance Transformclick-based promptingautoPET challengehuman-in-the-loop segmentation

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.

The reading

The paper is trying to establish that, in an interactive human-in-the-loop setting for whole-body PET/CT lesion segmentation, the way a user click is encoded into the network matters more than the choice of promptable architecture. Building on a strong fully automated baseline, the authors add two channels for foreground and background clicks and compare Gaussian-kernel encodings against Euclidean Distance Transform (EDT) encodings. They report that EDT encodings consistently improve Dice and reduce false-positive and false-negative volumes across fivefold cross-validation, especially when combined with online click simulation and a custom point sampler. If the claim holds, interactive PET/CT segmentation can be made more reliable without redesigning the segmentation model, and the gains translate into fewer manual corrections.

Core claim

The central claim is that a promptable U-Net-style PET/CT lesion segmenter conditioned by Euclidean Distance Transform (EDT) click channels beats the same model conditioned by Gaussian-kernel click channels. The paper shows this through five-fold cross-validation on the official autoPET/CT IV Task 1 split, where the best EDT variant (Size 2, custom points) achieves the highest Dice and the lowest false-positive and false-negative volumes among the single models. The authors further claim that training-time simulation of clicks is most robust when the official challenge sampler is mixed with a custom sampler that places positive clicks anywhere inside the lesion, favoring the lesion core, and

What carries the argument

The central object is the Euclidean Distance Transform (EDT) click encoding: each user click is expanded into a channel whose values are scaled distances from the clicked voxel, rather than a normalized Gaussian bump. The paper argues EDT wins because Gaussian kernels normalized to unit volume produce very low voxel intensities that the network effectively ignores. The second mechanism is the online prompt simulator, which, after data augmentation, draws 0-10 foreground and 0-10 background clicks per training patch from an 80/20 mixture of the official challenge sampler and a custom sampler, exposing the network to a wider variety of click locations than the official distribution alone.

Load-bearing premise

The gains rest on the assumption that clicks simulated during training resemble the clicks users will actually make at test time; if clinicians click in different places or with different frequency, the reported advantage of EDT encodings and custom point sampling may not transfer.

What would settle it

A direct comparison on the official test set, or in a study where radiologists provide real clicks, in which Gaussian-kernel encodings match or exceed EDT encodings on Dice or false-positive volume under the same five-fold protocol would refute the central claim; the paper already specifies the evaluation setup needed to run that check.

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

If this is right

Interactive PET/CT segmenters can be built from existing automated U-Net pipelines by adding two prompt channels, leaving architecture, loss, and hyperparameters untouched.
Reported reductions in false-positive and false-negative volumes imply that a few user clicks can yield masks that need less manual cleanup in clinical workflows.
The online simulation recipe offers a way to train promptable models when real clinician clicks are scarce, by generating realistic prompt distributions on the fly.
The ensemble result suggests combining a strong in-domain baseline with a model trained on additional external data delivers the most robust interactive performance.

Where Pith is reading between the lines

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

The EDT-versus-Gaussian outcome is measured on one U-Net encoder; transformer-based promptable models may respond differently to low-intensity Gaussian prompt channels, so the representation result should not be assumed universal without a similar ablation.
Real clinicians may click lesion boundaries or uncertain regions more often than the simulated sampler does; a reader study with actual clicks would test whether the custom point-sampling advantage survives real-world prompting behavior.
The logarithmic distribution over click counts implies diminishing returns from additional clicks; an explicit click-budget analysis could give users a stopping rule during interactive refinement.
Because the external training data spans different tracers, the EDT conditioning could be helping the model combine prompt information with tracer-specific uptake patterns, a hypothesis worth isolating in a per-tracer analysis.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

A useful ablation of click encodings for interactive PET/CT segmentation, but the headline ensemble result is not actually reported.

read the letter

You should read this as a challenge report, not a full methods paper. The group took their winning autoPET III pipeline and bolted on promptable click channels using the nnInteractive idea. That is fine, and they are honest about it. What is actually new is the systematic comparison of Euclidean distance transform (EDT) versus Gaussian click encodings for whole-body PET/CT, plus a custom click-sampling strategy. That comparison is useful and reasonably done on an official cross-validation split, and they released code. Credit where due: the experiments are transparent, the provenance is clear, and the improvement in last-click Dice (roughly 68 to 76) is a real signal that the prompt design matters.

The soft spots are proportionate but real. Most importantly, the abstract and conclusion claim that their final ensemble 'achieves the strongest cross-validation performance,' but Table 1 contains no ensemble row. There is no Dice, FPvol, or FNvol for the submitted model. The stress-test note is correct: the two ensembled models are best on different metrics, so you cannot infer the ensemble's performance. That is a load-bearing gap because the ensemble is the actual submission. Second, 'consistently outperformed' is too strong. Table 1 shows the EDT models beat the Gaussian kernels on Dice and FNvol at the last click, but some EDT variants have worse FPvol AUC than the best Gaussian (e.g., EDT Size2 has FPvol AUC 26.31 vs. sigma=0.25's 25.33). The trend is favorable, not universal. Third, there are no error bars or significance tests, and variant selection happens on the same cross-validation that supports the claim, so the margin between the best EDT and best Gaussian could be noise. Fourth, the custom point sampler is evaluated on the official validation prompts that were used to tune the mix ratio; that is mild overfitting to the validation set, and the paper does not dwell on it.

The citation pattern is expected—they build on their own nnInteractive, LesionLocator, and autoPET III work—so I would not call it a flaw, especially since code is public. The paper is honest about the inheritance.

This is not a fundamentally new mechanism; it is an application and ablation for a specific challenge. That is fine for a workshop or short paper. The right peer-review outcome is major revision: add the ensemble's numbers, add error bars or at least paired comparisons, soften the 'consistently' claim, and state how sensitive the custom point sampler is to the 80/20 mix. The work deserves referee time, not a desk reject, because the empirical question is well-posed and the code is available.

Referee Report

3 major / 4 minor

Summary. The paper presents the authors' submission to the autoPET/CT IV challenge Task 1 (interactive lesion segmentation in whole-body PET/CT). Building on the winning autoPET III nnU-Net pipeline, the authors add two input channels that encode simulated foreground/background clicks, and compare Gaussian-kernel and Euclidean Distance Transform (EDT) encodings. They also propose online simulation of user interactions and a custom point-sampling strategy. Five-fold cross-validation on the official split (Table 1) compares the baseline, several Gaussian variants, EDT variants, and combinations with custom points and additional external data. The abstract and conclusion claim that an ensemble of the two best EDT Size2 models (with and without additional data) achieves the strongest cross-validation performance, while reducing false positives and false negatives. The paper reports code and algorithm details (Table 2).

Significance. If substantiated, the paper provides a practical recipe for turning a strong fully-automatic PET/CT lesion segmentation model into a promptable interactive system, with the specific finding that EDT encodings outperform Gaussian kernels for click conditioning. This is a useful, well-scoped contribution to a challenge and to the broader promptable-segmentation literature. Strengths include the use of a strong, publicly known baseline, a systematic comparison of click-encoding choices, a transparent cross-validation protocol, and the release of code. However, the central claim about the final ensemble is not quantitatively supported, and the reported metrics lack uncertainty quantification, which limits the strength of the conclusions until those gaps are addressed. The work is a credible challenge-report contribution but does not yet provide the evidence needed for a broader methodological claim.

major comments (3)

[Abstract, §3.1, §4] The abstract and conclusion assert that the ensemble of the EDT Size2 model with custom points and the same model trained with additional data 'achieves the strongest cross-validation performance' and reduces false positives and false negatives compared to baseline models. However, §3.1 does not report the ensemble's Dice, FPvol, or FNvol anywhere; Table 1 lists only individual models. The two ensemble components are best on different metrics ('EDT Size2, custom points' has lowest FPvol of 8.05; 'EDT Size2, more data, custom points' has highest Dice of 76.35), so the ensemble's performance is not inferable from the table. Since the ensemble is the actual submitted method, its quantitative evaluation is essential. Please add the ensemble's metrics computed with the same official evaluation protocol.
[Table 1, §3] All results are reported as point estimates without error bars, confidence intervals, or significance tests. Multiple model variants are compared on the same fixed cross-validation splits and the best-performing variants are then selected based on these metrics, introducing a selection-bias risk. For example, the gap between the best EDT variant (Dice 76.35) and the best Gaussian variant (Dice 74.59) is about 1.8 Dice points, and the FPvol differences are smaller; without per-fold variability or paired testing, the claim that EDT 'consistently outperforms' Gaussian kernels is not fully supported. Please report per-fold results, standard deviations, and/or paired statistical tests for the key comparisons in Table 1.
[§3, Table 1] The text states that 'the model with Size 2 and custom points achieves strong performance, reducing FPvol and FNvol compared to other configurations.' This is contradicted by Table 1: while 'EDT Size2, custom points' has the lowest FPvol (8.05), its FNvol (AUC 15.11, Last 4.19) is higher than that of plain 'EDT Size2' (14.46, 3.86) and 'EDT Size4' (13.67, 3.66). Similarly, the claim that the 'custom point sampling strategy provides better generalization to realistic click distributions' is supported only partially (FPvol improves, Dice is similar, FNvol worsens). Please correct the description or provide additional evidence, such as an analysis of the official validation prompt distribution, to support the claim.

minor comments (4)

[Figure 1 caption] Typography: 'Exemplary' should be 'Example' or 'Exemplar' (the current spelling is not standard).
[Table 1] The '6,14' in the first Gaussian row appears to be a comma-decimal error; should be '6.14' for consistency with the rest of the table.
[§3, Table 1] The table header 'Dice↑ FPvol↓ FNvol↓' with subcolumns 'AUC Last' is not fully explained. Please define what 'AUC' and 'Last' refer to (e.g., area under the curve across the number of clicks evaluated at 0, 3, 7, 10, and the value at 10 clicks, respectively), preferably in the caption or in the text.
[§3, Table 1] The phrase 'AUC calculated using 0, 3, 7, and 10 clicks' is ambiguous: is this the average of the metric across these click counts, or an area under the curve computed across these points? Please clarify.

Circularity Check

0 steps flagged

No circular derivation: empirical ablations and official validation, though some claims lack reported support.

full rationale

The paper's central claims are empirical comparisons: EDT vs Gaussian click encodings, and variants of prompt simulation/training data. These are supported by five-fold cross-validation on the official challenge splits and the official evaluation implementation (Section 3, Table 1). There is no equation in which an output quantity is defined in terms of itself, no fitted parameter is renamed as a prediction, and no uniqueness theorem is invoked. The EDT representation is said to be 'inspired by nnInteractive' (Section 2), but its superiority is tested against Gaussian kernels rather than assumed from the citation. The model builds on the authors' previous autoPET III pipeline, nnInteractive, and LesionLocator; these are prior independent works, and the autoPET III baseline is externally validated from the earlier challenge. The custom point sampler is trained with an 80/20 mixture and then evaluated on the official precomputed validation prompts, so the comparison is a held-out ablation rather than a self-fulfilling prediction, although the phrase 'better generalization to realistic click distributions' is somewhat subjective. The most notable gap is that the abstract's claim that the ensemble 'achieves the strongest cross-validation performance' is not backed by any row in Table 1; Section 3.1 describes the ensemble but reports no Dice, FPvol, or FNvol value for it, and no error bars or paired tests are provided. This is a missing-evidence/support problem, not a circularity problem. No specific reduction of a result to its inputs by construction was found, so the circularity score is 0.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The paper relies on no invented entities and makes no theoretical derivation. Its load-bearing choices are empirical: the EDT encoding size and prompt-sampling distributions are selected from the same validation set used to evaluate them, and the external-data result assumes cross-dataset label compatibility. The central claim therefore rests on the assumption that the official click simulator and validation folds represent real interactive use.

free parameters (5)

EDT scaling factor (size) = 2
Prompt-encoding constant selected after comparing sizes 2, 3, and 4 on the official validation split; central to the claim that EDT outperforms Gaussian encodings.
Gaussian sigma test set = 0.1, 0.25, 0.5, 0.75, 1, 2, 3
Hand-chosen range of Gaussian kernel widths used to represent the alternative prompt encoding; the superiority claim is relative to this range.
Simulated prompt-count distribution = logarithmic, 0-10 per class
Number of positive and negative clicks per training patch is sampled with logarithmic probability favoring few points; a design choice for online interaction simulation.
Mix ratio between official and custom prompt simulators = 80/20
Online prompts are sampled 80% from official challenge code and 20% from the custom implementation; chosen by hand to balance realism and variability.
Ensemble composition = 5-fold ensemble of EDT Size2 with custom points, with and without extra data
The final submission combines two selected models; composition was determined from cross-validation results on the official splits.

axioms (4)

domain assumption Simulated user clicks (official challenge simulator plus 20% custom sampling) are representative of real clinician prompts at test time.
Invoked in Section 2 under 'Simulating user interaction'; if false, all interactive improvements reported in Table 1 may not transfer to actual use.
domain assumption Cross-validation on the official challenge training splits is a valid predictor of final challenge test performance.
The paper selects models and claims 'best cross-validation performance' based on official folds (Section 3), without reporting test-set metrics.
domain assumption External datasets (DEEP-PSMA, TCGA-LUAD, NSCLC-Radiogenomics) are label-compatible with the autoPET lesion definition and can be pooled with challenge data.
The 'more data' variant relies on pooling these datasets; no label-adjudication or domain-shift analysis is provided.
domain assumption Adding two prompt channels does not break the assumptions of the nnU-Net normalization, resampling, or loss used in the baseline.
The paper keeps architecture, loss, and hyperparameters unchanged (Section 2); this presumes that the extra channels integrate cleanly with the pretrained pipeline.

reviewed 2026-08-05 · how reviews work

0 comments

Cite this review

Pith. "Pith review of Towards Interactive Lesion Segmentation in Whole-Body PET/CT with Promptable Models." pith.science (2026). https://pith.science/paper/LK4BDFFQ

@misc{pith2026250821680,
  author       = {Pith},
  title        = {Pith review of: Towards Interactive Lesion Segmentation in Whole-Body PET/CT with Promptable Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LK4BDFFQ}},
  note         = {Machine review of arXiv:2508.21680}
}

Share X Bluesky LinkedIn Reddit HN

read the original abstract

Whole-body PET/CT is a cornerstone of oncological imaging, yet accurate lesion segmentation remains challenging due to tracer heterogeneity, physiological uptake, and multi-center variability. While fully automated methods have advanced substantially, clinical practice benefits from approaches that keep humans in the loop to efficiently refine predicted masks. The autoPET/CT IV challenge addresses this need by introducing interactive segmentation tasks based on simulated user prompts. In this work, we present our submission to Task 1. Building on the winning autoPET III nnU-Net pipeline, we extend the framework with promptable capabilities by encoding user-provided foreground and background clicks as additional input channels. We systematically investigate representations for spatial prompts and demonstrate that Euclidean Distance Transform (EDT) encodings consistently outperform Gaussian kernels. Furthermore, we propose online simulation of user interactions and a custom point sampling strategy to improve robustness under realistic prompting conditions. Our ensemble of EDT-based models, trained with and without external data, achieves the strongest cross-validation performance, reducing both false positives and false negatives compared to baseline models. These results highlight the potential of promptable models to enable efficient, user-guided segmentation workflows in multi-tracer, multi-center PET/CT. Code is publicly available at https://github.com/MIC-DKFZ/autoPET-interactive

Figures

Figures reproduced from arXiv: 2508.21680 by Fabian Isensee, Klaus H. Maier-Hein, Maximilian Rokuss, Yannick Kirchhoff.

discussion (0)

Reference graph

Works this paper leans on

13 extracted references · 7 canonical work pages

[1]

JGNIHEP5, https://doi.org/10.7937/K9/TCIA.2016.JGNIHEP5, data set

Albertina, B., Watson, M., Holback, C., Jarosz, R., Kirk, S., Lee, Y., Rieger- Christ, K., Lemmerman, J.: The cancer genome atlas lung adenocarcinoma col- lection (tcga-luad) (version 4) (2016).https://doi.org/10.7937/K9/TCIA.2016. JGNIHEP5, https://doi.org/10.7937/K9/TCIA.2016.JGNIHEP5, data set

work page doi:10.7937/k9/tcia.2016 2016
[2]

Scientific data5(1), 1–9 (2018)

Bakr, S., Gevaert, O., Echegaray, S., Ayers, K., Zhou, M., Shafiq, M., Zheng, H., Benson, J.A., Zhang, W., Leung, A.N., et al.: A radiogenomic dataset of non-small cell lung cancer. Scientific data5(1), 1–9 (2018)

work page 2018
[3]

Ferdinandus, J., Violet, J., Sandhu, S., Hicks, R.J., Ravi Kumar, A.S., Iravani, A., Kong, G., Akhurst, T., Thang, S.P., Murphy, D.G., et al.: Prognostic biomarkers in men with metastatic castration-resistant prostate cancer receiving [177lu]-psma-

work page
[4]

Gatidis, S., Kuestner, T.: A whole-body fdg-pet/ct dataset with manually an- notated tumor lesions (fdg-pet-ct-lesions).https://doi.org/10.7937/GKR0-XV29 (2022), the Cancer Imaging Archive

doi:10.7937/gkr0-xv29 2022
[5]

Nature methods 18(2), 203–211 (2021)

Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods 18(2), 203–211 (2021). https://doi.org/https://doi.org/10. 1038/s41592-020-01008-z

work page 2021
[6]

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 (2025),https: //arxiv.org/abs/2503.08373

Pith/arXiv arXiv 2025
[7]

arXiv preprint arXiv:2404.09556 (2024)

Isensee, F., Wald, T., Ulrich, C., Baumgartner, M., Roy, S., Maier-Hein, K., Jaeger, P.F.: nnU-Net revisited: A call for rigorous validation in 3d medical image segmen- tation. arXiv preprint arXiv:2404.09556 (2024)

Pith/arXiv arXiv 2024
[8]

https://doi.org/10.7937/r7ep-3x37 (2024), the Cancer Imaging Archive

Jeblick, K., et al.: A whole-body psma-pet/ct dataset with manually annotated tu- mor lesions (psma-pet-ct-lesions). https://doi.org/10.7937/r7ep-3x37 (2024), the Cancer Imaging Archive

doi:10.7937/r7ep-3x37 2024
[9]

arXiv:2304.02643 (2023)

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. arXiv:2304.02643 (2023)

Pith/arXiv arXiv 2023
[10]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Rokuss, M., Kirchhoff, Y., Akbal, S., Kovacs, B., Roy, S., Ulrich, C., Wald, T., Rotkopf, L.T., Schlemmer, H.P., Maier-Hein, K.: Lesionlocator: Zero-shot universal tumor segmentation and tracking in 3d whole-body imaging. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 30872–30885 (June 2025)

work page 2025
[11]

Rokuss, M., Kovacs, B., Kirchhoff, Y., Xiao, S., Ulrich, C., Maier-Hein, K.H., Isensee, F.: From fdg to psma: A hitchhiker’s guide to multitracer, multicenter le- sion segmentation in pet/ct imaging (2024),https://arxiv.org/abs/2409.09478

Pith/arXiv arXiv 2024
[12]

In: Inter- national Conference on Medical Image Computing and Computer-Assisted Inter- vention

Ulrich, C., Isensee, F., Wald, T., Zenk, M., Baumgartner, M., Maier-Hein, K.H.: Multitalent: A multi-dataset approach to medical image segmentation. In: Inter- national Conference on Medical Image Computing and Computer-Assisted Inter- vention. pp. 648–658. Springer (2023) Interactive Lesion Segmentation in PET/CT 7 Table 2: Algorithm details T eam name a...

work page 2023
[617]

European journal of nuclear medicine and molecular imaging47(10), 2322– 2327 (2020)

work page 2020

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.

[1] [1]

JGNIHEP5, https://doi.org/10.7937/K9/TCIA.2016.JGNIHEP5, data set

Albertina, B., Watson, M., Holback, C., Jarosz, R., Kirk, S., Lee, Y., Rieger- Christ, K., Lemmerman, J.: The cancer genome atlas lung adenocarcinoma col- lection (tcga-luad) (version 4) (2016).https://doi.org/10.7937/K9/TCIA.2016. JGNIHEP5, https://doi.org/10.7937/K9/TCIA.2016.JGNIHEP5, data set

work page doi:10.7937/k9/tcia.2016 2016

[2] [2]

Scientific data5(1), 1–9 (2018)

Bakr, S., Gevaert, O., Echegaray, S., Ayers, K., Zhou, M., Shafiq, M., Zheng, H., Benson, J.A., Zhang, W., Leung, A.N., et al.: A radiogenomic dataset of non-small cell lung cancer. Scientific data5(1), 1–9 (2018)

work page 2018

[3] [3]

Ferdinandus, J., Violet, J., Sandhu, S., Hicks, R.J., Ravi Kumar, A.S., Iravani, A., Kong, G., Akhurst, T., Thang, S.P., Murphy, D.G., et al.: Prognostic biomarkers in men with metastatic castration-resistant prostate cancer receiving [177lu]-psma-

work page

[4] [4]

Gatidis, S., Kuestner, T.: A whole-body fdg-pet/ct dataset with manually an- notated tumor lesions (fdg-pet-ct-lesions).https://doi.org/10.7937/GKR0-XV29 (2022), the Cancer Imaging Archive

doi:10.7937/gkr0-xv29 2022

[5] [5]

Nature methods 18(2), 203–211 (2021)

Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods 18(2), 203–211 (2021). https://doi.org/https://doi.org/10. 1038/s41592-020-01008-z

work page 2021

[6] [6]

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 (2025),https: //arxiv.org/abs/2503.08373

Pith/arXiv arXiv 2025

[7] [7]

arXiv preprint arXiv:2404.09556 (2024)

Isensee, F., Wald, T., Ulrich, C., Baumgartner, M., Roy, S., Maier-Hein, K., Jaeger, P.F.: nnU-Net revisited: A call for rigorous validation in 3d medical image segmen- tation. arXiv preprint arXiv:2404.09556 (2024)

Pith/arXiv arXiv 2024

[8] [8]

https://doi.org/10.7937/r7ep-3x37 (2024), the Cancer Imaging Archive

Jeblick, K., et al.: A whole-body psma-pet/ct dataset with manually annotated tu- mor lesions (psma-pet-ct-lesions). https://doi.org/10.7937/r7ep-3x37 (2024), the Cancer Imaging Archive

doi:10.7937/r7ep-3x37 2024

[9] [9]

arXiv:2304.02643 (2023)

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. arXiv:2304.02643 (2023)

Pith/arXiv arXiv 2023

[10] [10]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Rokuss, M., Kirchhoff, Y., Akbal, S., Kovacs, B., Roy, S., Ulrich, C., Wald, T., Rotkopf, L.T., Schlemmer, H.P., Maier-Hein, K.: Lesionlocator: Zero-shot universal tumor segmentation and tracking in 3d whole-body imaging. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 30872–30885 (June 2025)

work page 2025

[11] [11]

Rokuss, M., Kovacs, B., Kirchhoff, Y., Xiao, S., Ulrich, C., Maier-Hein, K.H., Isensee, F.: From fdg to psma: A hitchhiker’s guide to multitracer, multicenter le- sion segmentation in pet/ct imaging (2024),https://arxiv.org/abs/2409.09478

Pith/arXiv arXiv 2024

[12] [12]

In: Inter- national Conference on Medical Image Computing and Computer-Assisted Inter- vention

Ulrich, C., Isensee, F., Wald, T., Zenk, M., Baumgartner, M., Maier-Hein, K.H.: Multitalent: A multi-dataset approach to medical image segmentation. In: Inter- national Conference on Medical Image Computing and Computer-Assisted Inter- vention. pp. 648–658. Springer (2023) Interactive Lesion Segmentation in PET/CT 7 Table 2: Algorithm details T eam name a...

work page 2023

[13] [617]

European journal of nuclear medicine and molecular imaging47(10), 2322– 2327 (2020)

work page 2020

Towards Interactive Lesion Segmentation in Whole-Body PET/CT with Promptable Models

The reading

Core claim §

What carries the argument

Load-bearing premise §

What would settle it §

If this is right

Where Pith is reading between the lines

discussion (0)

Core claim

Load-bearing premise

What would settle it