pith. sign in

arxiv: 2606.30108 · v1 · pith:AHVLWSYHnew · submitted 2026-06-29 · 💻 cs.CV

LETT-NeXt: A Lightweight RECIST-Guided Model for 3D CT Lesion Segmentation

Sebastian Aas , Elias Stenhede , Arian Ranjbar This is my paper

Pith reviewed 2026-06-30 06:48 UTC · model grok-4.3

classification 💻 cs.CV
keywords RECIST-guided segmentation3D CT lesion segmentationlightweight modelprompt channelsMedNeXttumor segmentationmedical imagingpan-cancer CT
0
0 comments X

The pith

LETT-NeXt predicts 3D lesion masks from CT by adding RECIST line and endpoints as two prompt channels to a centered crop.

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

The paper presents LETT-NeXt as a way to convert standard 2D RECIST diameter measurements into full 3D segmentations of lesions in CT volumes. It centers a regional crop on the RECIST markers, adds two dedicated prompt channels that mark the line and its endpoints, and feeds the combined input into a compact MedNeXt-v2 encoder-decoder. The resulting masks are refined with prompt-aware component selection and adaptive AutoZoom inference. On public validation data the model reaches a Dice score of 79.4 and on the hidden test set it reaches 73.9, with CPU inference completing in under seven seconds per case. The approach is offered for the CVPR 2026 pan-cancer CT segmentation challenge.

Core claim

LETT-NeXt extracts a RECIST-centered regional crop from the CT volume, encodes the RECIST line and endpoints as two additional prompt channels, concatenates them with the CT input, and passes the result through a compact MedNeXt-v2 encoder-decoder to output the 3D lesion mask; prompt-aware component selection and adaptive AutoZoom inference then refine the output, yielding DSC 79.4 ± 10.1 and NSD 72.3 ± 16.2 on public validation and DSC 73.9, NSD 67.3 on the hidden test set.

What carries the argument

RECIST prompt channels: two extra input channels that encode the RECIST line and endpoints, concatenated with the CT volume inside a RECIST-centered crop to steer the lightweight segmentation network.

If this is right

  • 3D lesion volumes become available from existing routine RECIST annotations without new manual contouring.
  • Inference runs on standard CPU hardware in 6.9 seconds per case with peak memory of 3.6 GB.
  • Prompt-aware component selection removes extraneous regions that the network might otherwise include.
  • Adaptive AutoZoom inference adjusts to lesions of different sizes within the same model.

Where Pith is reading between the lines

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

  • The same prompt-channel idea could be tested on other 2D clinical measurements such as caliper distances or long-axis lengths.
  • Removing the RECIST prompts entirely and retraining might reveal how much the network relies on image texture versus explicit location cues.
  • The lightweight footprint suggests the architecture could serve as a starting point for on-device or edge deployment in clinical workflows.

Load-bearing premise

Marking the RECIST line and endpoints as two prompt channels inside a centered crop supplies enough guidance for the network to recover accurate 3D lesion masks from the CT volume alone.

What would settle it

Measure Dice and NSD on the hidden test set after removing the two RECIST prompt channels while keeping every other part of the pipeline unchanged; a drop below 70 would indicate the prompts are required for the reported performance.

Figures

Figures reproduced from arXiv: 2606.30108 by Arian Ranjbar, Elias Stenhede, Sebastian Aas.

Figure 1
Figure 1. Figure 1: LETT-NeXt converts a RECIST marker into a volumetric lesion mask. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the MedNeXt-v2 f32 architecture used in LETT-NeXt. The [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative validation examples across score percentiles. Each row shows [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative examples of RECIST-guided AutoZoom on the validation set. [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Core MedNeXt-v2 building blocks used in LETT-NeXt. The main block [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
read the original abstract

RECIST diameter measurements are widely used for tumor response assessment, but they provide only a limited 2D description of lesion extent. We present LETT-NeXt, a lightweight RECIST-guided model that predicts 3D lesion masks from CT volumes and RECIST markers for the CVPR 2026 Foundation Models for Pan-cancer Segmentation in CT Images competition. LETT-NeXt extracts a RECIST-centered regional crop, encodes the RECIST line and endpoints as two prompt channels, and concatenates them with the CT input. A compact MedNeXt-v2 encoder--decoder predicts the lesion mask, followed by prompt-aware component selection and adaptive AutoZoom inference. On the public validation set, LETT-NeXt achieved a Dice Similarity Coefficient (DSC) of 79.4 $\pm$ 10.1 and a Normalized Surface Dice (NSD) of 72.3 $\pm$ 16.2. On the hidden test set, it achieved a DSC of 73.9 and an NSD of 67.3, corresponding to a challenge score of 70.6\%. On the public validation mirror, LETT-NeXt completed CPU inference in 6.9 $\pm$ 3.0 s per case with a peak memory use of 3.6 GB. Code is available at github.com/Ahus-AIM/lett-next.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The manuscript introduces LETT-NeXt, a lightweight RECIST-guided model for 3D CT lesion segmentation in the CVPR 2026 competition. It extracts a RECIST-centered regional crop from the CT volume, encodes the RECIST line and endpoints as two additional prompt channels, concatenates them with the CT input, and feeds the result to a compact MedNeXt-v2 encoder-decoder. This is followed by prompt-aware component selection and adaptive AutoZoom inference. The paper reports DSC of 79.4 ± 10.1 and NSD of 72.3 ± 16.2 on the public validation set, DSC of 73.9 and NSD of 67.3 on the hidden test set (challenge score 70.6%), plus CPU inference time of 6.9 ± 3.0 s per case and peak memory of 3.6 GB. Code is released at github.com/Ahus-AIM/lett-next.

Significance. If the reported performance holds under scrutiny, the work supplies a practical, memory-efficient pipeline that converts 2D RECIST annotations into 3D lesion masks for pan-cancer CT segmentation. The explicit release of code and the inclusion of standardized efficiency metrics on CPU constitute concrete strengths that enhance reproducibility and deployment value.

major comments (1)
  1. [Model description and experimental results] The central claim that encoding the RECIST line and endpoints as two prompt channels guides accurate 3D mask prediction rests on an untested assumption. No ablation is provided that removes or zeros these prompt channels while retaining the RECIST-centered crop, MedNeXt-v2 backbone, component selection, and AutoZoom (see the model description in the abstract and the implied methods). Without this isolation, it remains possible that the regional crop and base architecture alone account for the reported DSC/NSD values.
minor comments (2)
  1. [Abstract] Training procedure, loss functions, optimizer settings, and data augmentation details are absent, limiting assessment of the experimental protocol.
  2. [Results] The reported standard deviations (±10.1 DSC, ±16.2 NSD) are large; a brief discussion of failure cases or case-level variability would strengthen the results section.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for highlighting the need for an ablation to isolate the contribution of the RECIST prompt channels. We address the comment below and will revise the manuscript to include the requested experiment.

read point-by-point responses

  1. Referee: The central claim that encoding the RECIST line and endpoints as two prompt channels guides accurate 3D mask prediction rests on an untested assumption. No ablation is provided that removes or zeros these prompt channels while retaining the RECIST-centered crop, MedNeXt-v2 backbone, component selection, and AutoZoom (see the model description in the abstract and the implied methods). Without this isolation, it remains possible that the regional crop and base architecture alone account for the reported DSC/NSD values.

    Authors: We agree that the current manuscript lacks an ablation that removes or zeros the prompt channels while retaining the RECIST-centered crop, MedNeXt-v2 backbone, component selection, and AutoZoom. This is a valid observation. In the revised manuscript we will add this ablation study, reporting the resulting DSC and NSD on the public validation set to quantify the incremental contribution of the prompt channels. revision: yes

Circularity Check

0 steps flagged

No circularity detected; empirical results on held-out partitions

full rationale

The paper reports DSC and NSD metrics on distinct public validation and hidden test sets after describing a RECIST-prompt architecture. No equations, fitted parameters, or self-citations are present that would reduce the performance numbers to a definitional identity or construction from the inputs. The method description (crop + prompt channels + MedNeXt-v2 + post-processing) stands as an independent design choice whose outputs are evaluated externally.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no explicit free parameters, axioms, or invented entities; the model description implies standard neural-network training assumptions that are not enumerated.

pith-pipeline@v0.9.1-grok · 5788 in / 1199 out tokens · 23128 ms · 2026-06-30T06:48:58.063844+00:00 · methodology

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

31 extracted references · 18 canonical work pages · 7 internal anchors

  1. [1]

    FLARE-MedFM/FLARE-Task1-PancancerRECIST-to-3D·Datasets at Hugging Face

  2. [2]

    Lite ENSAM: a lightweight cancer segmentation model for 3D Computed Tomography, November

    Agnar Martin Bjornstad, Elias Stenhede, and Arian Ranjbar. Lite ENSAM: a lightweight cancer segmentation model for 3D Computed Tomography, November

  3. [3]

    arXiv:2511.01600 [cs] version: 1

    work page arXiv

  4. [4]

    SAM 3: Segment Anything with Concepts, March

    Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoubhik Debnath, Ronghang Hu, Didac Suris, Chaitanya Ryali, Kalyan Vasudev Alwala, Haitham Khedr, An- drew Huang, Jie Lei, Tengyu Ma, Baishan Guo, Arpit Kalla, Markus Marks, Joseph Greer, Meng Wang, Peize Sun, Roman R¨ adle, Triantafyllos Afouras, Effrosyni Mavroudi, Katherine Xu, Tsung-Han Wu, Yu Zhou, Lili...

  5. [5]

    SAM 3: Segment Anything with Concepts

    arXiv:2511.16719 [cs.CV]. LETT-NeXt for CT Lesion Segmentation 17

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [6]

    SegVol: Universal and Inter- active Volumetric Medical Image Segmentation, February 2025

    Yuxin Du, Fan Bai, Tiejun Huang, and Bo Zhao. SegVol: Universal and Inter- active Volumetric Medical Image Segmentation, February 2025. arXiv:2311.13385 [cs.CV]

    work page arXiv 2025

  7. [7]

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

    2009

  8. [8]

    Hayes, M.C

    S.A. Hayes, M.C. Pietanza, D. O’Driscoll, J. Zheng, C.S. Moskowitz, M.G. Kris, and M.S. Ginsberg. Comparison of CT volumetric measurement with RECIST response in patients with lung cancer.European Journal of Radiology, 85(3):524– 533, March 2016

    2016

  9. [9]

    VISTA3D: A Unified Segmentation Foundation Model For 3D Medical Imaging, November 2024

    Yufan He, Pengfei Guo, Yucheng Tang, Andriy Myronenko, Vishwesh Nath, Ziyue Xu, Dong Yang, Can Zhao, Benjamin Simon, Mason Belue, Stephanie Harmon, Baris Turkbey, Daguang Xu, and Wenqi Li. VISTA3D: A Unified Segmentation Foundation Model For 3D Medical Imaging, November 2024. arXiv:2406.05285 [cs]

    work page arXiv 2024

  10. [10]

    Huang, et al., STU-Net: Scalable and transferable medical im- age segmentation models empowered by large-scale supervised pre- trainingArXiv:2304.06716

    Ziyan Huang, Haoyu Wang, Zhongying Deng, Jin Ye, Yanzhou Su, Hui Sun, Jun- jun He, Yun Gu, Lixu Gu, Shaoting Zhang, and Yu Qiao. STU-Net: Scalable and Transferable Medical Image Segmentation Models Empowered by Large-Scale Supervised Pre-training, April 2023. arXiv:2304.06716 [cs]

    work page arXiv 2023

  11. [11]

    Mertens, Felix J

    Hartmut H¨ antze, Lina Xu, Christian J. Mertens, Felix J. Dorfner, Leonhard Donle, Felix Busch, Avan Kader, Sebastian Ziegelmayer, Nadine Bayerl, Nas- sir Navab, Daniel Rueckert, Julia Schnabel, Hugo JWL Aerts, Daniel Truhn, Fabian Bamberg, Jakob Weiß, Christopher L. Schlett, Steffen Ringhof, Thoralf Niendorf, Tobias Pischon, Hans-Ulrich Kauczor, Tobias N...

  12. [12]

    arXiv:2405.06463 [eess]

    work page arXiv

  13. [13]

    nnU-Net: Self-adapting Framework for U-Net-Based Medical Image Segmentation

    Fabian Isensee, Jens Petersen, Andre Klein, David Zimmerer, Paul F. Jaeger, Si- mon Kohl, Jakob Wasserthal, Gregor Koehler, Tobias Norajitra, Sebastian Wirk- ert, and Klaus H. Maier-Hein. nnU-Net: Self-adapting Framework for U-Net-Based Medical Image Segmentation, September 2018. arXiv:1809.10486 [cs]

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  14. [14]

    nnInteractive: Redefining 3D Promptable Segmentation, March 2025

    Fabian Isensee, Maximilian Rokuss, Lars Kr¨ amer, Stefan Dinkelacker, Ashis Ravin- dran, Florian Stritzke, Benjamin Hamm, Tassilo Wald, Moritz Langenberg, Con- stantin Ulrich, Jonathan Deissler, Ralf Floca, and Klaus Maier-Hein. nnInteractive: Redefining 3D Promptable Segmentation, March 2025

    2025

  15. [15]

    Segment Anything

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Pi- otr Doll´ ar, and Ross Girshick. Segment Anything, April 2023. arXiv:2304.02643 [cs.CV]

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  16. [16]

    Wenxuan Li, Xinze Zhou, Qi Chen, Tianyu Lin, Pedro R. A. S. Bassi, Szymon Plotka, Jaroslaw B. Cwikla, Xiaoxi Chen, Chen Ye, Zheren Zhu, Kai Ding, Heng Li, Kang Wang, Yang Yang, Yucheng Tang, Daguang Xu, Alan L. Yuille, and Zong- wei Zhou. PanTS: The Pancreatic Tumor Segmentation Dataset, 2025. Version Number: 1

    2025

  17. [17]

    2022, arXiv e-prints, arXiv:2201.03545, doi: 10.48550/arXiv.2201.03545 LSST Dark Energy Science Collaboration, Aubourg, E.,

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A ConvNet for the 2020s, March 2022. arXiv:2201.03545 [cs]. 18 Sebastian Aas, Elias Stenhede, and Arian Ranjbar

    work page arXiv 2022

  18. [18]

    Segment Anything in Medical Images.Nature Communications, 15(1):654, January 2024

    Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment Anything in Medical Images.Nature Communications, 15(1):654, January 2024. arXiv:2304.12306 [eess.IV]

    work page arXiv 2024

  19. [19]

    Springer Nature Switzerland, Cham, 2026

    Jun Ma and Bo Wang, editors.Fast, Low-Resource, Accurate Robust Organ and Pan-cancer Segmentation: MICCAI Challenge, FLARE 2024, Held in Conjunction with MICCAI 2024, Marrakesh, Morocco, October 6, 2024, Proceedings, volume 15717 ofLecture Notes in Computer Science. Springer Nature Switzerland, Cham, 2026

    2024

  20. [20]

    A promptable CT foundation model for solid tumor evaluation.npj Precision Oncology, 9(1):121, April 2025

    L´ eo Machado, L´ eo Alberge, H´ el` ene Philippe, Elodie Ferreres, Julien Khlaut, Julie Dupuis, Korentin Le Floch, Denis Habip Gatenyo, Pascal Roux, Jules Gr´ egory, Maxime Ronot, Corentin Dancette, Tom Boeken, Daniel Tordjman, Pierre Manceron, and Paul H´ erent. A promptable CT foundation model for solid tumor evaluation.npj Precision Oncology, 9(1):121...

    2025

  21. [21]

    3D MRI brain tumor segmentation using autoencoder regularization

    Andriy Myronenko. 3D MRI brain tumor segmentation using autoencoder regu- larization, November 2018. arXiv:1810.11654 [cs]

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  22. [22]

    SAM 2: Segment Anything in Images and Videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R¨ adle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Doll´ ar, and Christoph Feichtenhofer. SAM 2: Segment Any- thing in Images and Videos, October 2024. arXiv:24...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  23. [23]

    U-Net: Convolutional Networks for Biomedical Image Segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional Net- works for Biomedical Image Segmentation, May 2015. arXiv:1505.04597 [cs]

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  24. [24]

    MedNeXt-v2: Scaling 3D ConvNeXts for Large-Scale Supervised Representation Learning in Medical Image Segmenta- tion, December 2025

    Saikat Roy, Yannick Kirchhoff, Constantin Ulrich, Maximillian Rokuss, Tassilo Wald, Fabian Isensee, and Klaus Maier-Hein. MedNeXt-v2: Scaling 3D ConvNeXts for Large-Scale Supervised Representation Learning in Medical Image Segmenta- tion, December 2025. arXiv:2512.17774 [eess]

    work page arXiv 2025

  25. [25]

    J¨ ager, and Klaus H

    Saikat Roy, Gregor Koehler, Constantin Ulrich, Michael Baumgartner, Jens Pe- tersen, Fabian Isensee, Paul F. J¨ ager, and Klaus H. Maier-Hein. MedNeXt: Transformer-Driven Scaling of ConvNets for Medical Image Segmentation. In Hayit Greenspan, Anant Madabhushi, Parvin Mousavi, Septimiu Salcudean, James Dun- can, Tanveer Syeda-Mahmood, and Russell Taylor, e...

    2023

  26. [26]

    Springer Nature Switzerland

  27. [27]

    ENSAM: an efficient foundation model for interactive segmentation of 3D medical images, September

    Elias Stenhede, Agnar Martin Bjørnstad, and Arian Ranjbar. ENSAM: an efficient foundation model for interactive segmentation of 3D medical images, September

  28. [28]

    arXiv:2509.15874 [cs]

    work page internal anchor Pith review arXiv

  29. [29]

    SAM-Med3D: Towards General-purpose Segmentation Models for Volumetric Medical Images, September 2024

    Haoyu Wang, Sizheng Guo, Jin Ye, Zhongying Deng, Junlong Cheng, Tianbin Li, Jianpin Chen, Yanzhou Su, Ziyan Huang, Yiqing Shen, Bin Fu, Shaoting Zhang, Junjun He, and Yu Qiao. SAM-Med3D: Towards General-purpose Segmentation Models for Volumetric Medical Images, September 2024. arXiv:2310.15161 [cs.CV]

    work page arXiv 2024

  30. [30]

    Meyer, Maurice Pradella, Daniel Hinck, Alexander W

    Jakob Wasserthal, Hanns-Christian Breit, Manfred T. Meyer, Maurice Pradella, Daniel Hinck, Alexander W. Sauter, Tobias Heye, Daniel Boll, Joshy Cyriac, Shan Yang, Michael Bach, and Martin Segeroth. TotalSegmentator: robust segmenta- tion of 104 anatomical structures in CT images.Radiology: Artificial Intelligence, 5(5):e230024, September 2023. arXiv:2208....

    work page arXiv 2023

  31. [31]

    Codabench: Flexible, Easy-to-Use and Reproducible Benchmarking Platform, February 2022

    Zhen Xu, Sergio Escalera, Isabelle Guyon, Adrien Pav˜ ao, Magali Richard, Wei- Wei Tu, Quanming Yao, and Huan Zhao. Codabench: Flexible, Easy-to-Use and Reproducible Benchmarking Platform, February 2022. arXiv:2110.05802 [cs.LG]

    work page arXiv 2022