pith. sign in

arxiv: 2606.15604 · v1 · pith:XSOQIAKMnew · submitted 2026-06-14 · 📡 eess.IV · cs.CV

Parameter-Efficient Adaptation of SAM 3 for Automated ITV Generation from 4DCT Images

Changwoo Song This is my paper

Pith reviewed 2026-06-27 04:30 UTC · model grok-4.3

classification 📡 eess.IV cs.CV
keywords 4DCTSAM 3LoRAITV generationmedical image segmentationtemporal filteringradiotherapyparameter-efficient fine-tuning
0
0 comments X

The pith

Adapting SAM 3 via LoRA plus temporal filtering automates ITV generation from 4DCT scans with seven annotated volumes.

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

The paper shows how low-rank adaptation aligns the Segment Anything Model 3 to thoracic CT data so that text prompts produce reliable segmentations across the respiratory cycle. Phase-coherent filtering then keeps only structures that persist through consecutive phases while discarding sporadic false positives. This combination reaches median Dice scores above 0.91 on pulmonary and cardiac targets and keeps more than 95 percent of full-data performance. Readers care because current ITV contouring is manual, discards motion information, and requires far more labeled examples.

Core claim

Parameter-efficient fine-tuning of SAM 3 with LoRA and hard negative mining aligns its text-prompted output to the medical domain; phase-wise predictions are then refined by temporal filtering that exploits the continuous periodic nature of respiration, so genuine anatomy remains contiguous across phases while transient artifacts are suppressed. The resulting contours achieve median Dice scores of 0.968 for pulmonary and 0.910 for cardiac structures with 95th-percentile Hausdorff distances of 0.998 mm and 2.931 mm, eliminate the severe false positives of zero-shot SAM 3, and retain over 95 percent of full-data accuracy when trained on only seven annotated volumes.

What carries the argument

LoRA-adapted SAM 3 segmentation followed by phase-coherent temporal filtering and spatial connectivity analysis.

Load-bearing premise

Genuine anatomy appears in contiguous blocks of phases while transient artifacts appear sporadically.

What would settle it

A 4DCT test set containing irregular or non-periodic motion patterns in which the temporal filtering step fails to raise Dice scores above those of the unadapted zero-shot model.

Figures

Figures reproduced from arXiv: 2606.15604 by Changwoo Song.

Figure 1
Figure 1. Figure 1: Overview of the proposed framework. SAM 3 is adapted via LoRA and prompted with target structure names. Phase-wise 3D predictions are filtered by res￾piratory phase coherence and spatial connectivity to produce a unified ITV. 2.2 Parameter-Efficient Fine-Tuning with LoRA SAM 3 [2], the latest iteration of the Segment Anything Model, employs a vision transformer (ViT) [4] backbone pretrained on large-scale … view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative comparison across three axial slices. Red: prediction; green: ground truth. (a) Input CT. (b) Zero-shot SAM 3. (c) LoRA. (d) LoRA + HNM. (e) Ground truth. Top to bottom: central, boundary, and distant cardiac slices. that the proposed adaptation (LoRA + HNM) reaches near-saturation perfor￾mance with remarkably few samples. This confirms that LoRA’s low-rank update preserves the pretrained repre… view at source ↗
Figure 3
Figure 3. Figure 3: Data efficiency: DSC (left) and performance retention (right) as training set size decreases. Both organs retain >95% of full-data accuracy with only 7 cases. annotated 3D CT volumes—with a DSC degradation of under 0.04 relative to 79-case training—and applying it directly to unseen 4DCT phases, our approach circumvents the two most cited barriers to clinical deployment of deep learning segmentation: data … view at source ↗
Figure 4
Figure 4. Figure 4: Spatiotemporal filtering on representative cardiac axial slices. (a) AIP (average of 10 phases). (b) Union ITV (red). (c) Refined ITV (green, Rmin=3). (d) Retained ITV (green) and removed noise (yellow). Top: cardiac boundary; bottom: distant slice. The primary limitations of this study are twofold: (1) the 4DCT evaluation relies on qualitative expert assessment due to the inherent clinical difficulty of e… view at source ↗
read the original abstract

Four-dimensional computed tomography (4DCT) captures the full respiratory cycle of thoracic anatomy, yet current Internal Target Volume contouring workflows process each phase in isolation, discarding temporal coherence and leaving contours vulnerable to phase-specific artifacts. We present a lightweight framework that applies parameter-efficient fine-tuning to the Segment Anything Model 3 (SAM 3) via low-rank adaptation (LoRA) to align its text-prompted segmentation with the medical domain using only seven annotated 3D CT volumes. Furthermore, the framework incorporates a hard negative mining strategy to improve boundary discrimination in low-contrast thoracic regions. At inference, phase-wise predictions are refined through phase-coherent temporal filtering and spatial connectivity analysis. Since respiratory motion is continuous and periodic, genuine anatomy appears in contiguous blocks of phases, whereas transient artifacts appear sporadically and are thus effectively suppressed. Experiments on pulmonary and cardiac structures yield median Dice scores of 0.968 and 0.910 with 95th-percentile Hausdorff distances of 0.998 mm and 2.931 mm, respectively. The proposed framework effectively eliminates the severe false-positive predictions inherent in the zero-shot inference of the unadapted SAM 3. With only seven annotated volumes, the framework retains over 95% of full-data accuracy, and the entire pipeline is trainable on a single consumer-grade GPU, demonstrating a scalable, data-efficient solution for adaptive radiotherapy.

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

2 major / 0 minor

Summary. The manuscript proposes a lightweight framework for automated Internal Target Volume (ITV) generation from 4DCT images by applying parameter-efficient fine-tuning (LoRA) to SAM 3 using only seven annotated 3D CT volumes, combined with hard negative mining during training. At inference, phase-wise predictions are refined via phase-coherent temporal filtering and spatial connectivity analysis, justified by the claim that genuine anatomy appears in contiguous phase blocks while transient artifacts are sporadic. Experiments on pulmonary and cardiac structures report median Dice scores of 0.968 and 0.910 with 95th-percentile Hausdorff distances of 0.998 mm and 2.931 mm, claiming elimination of zero-shot false positives and retention of over 95% of full-data accuracy on a single consumer GPU.

Significance. If validated, the work demonstrates a data-efficient adaptation of a foundation model for a clinically relevant radiotherapy task, highlighting the potential of LoRA for medical domain alignment with minimal annotations and low compute. The explicit use of 4D temporal coherence is a distinguishing element, though its contribution to the reported gains requires substantiation to confirm the framework's advantages over zero-shot or fully supervised baselines.

major comments (2)
  1. [Abstract] Abstract (inference stage): The headline performance metrics and the claim of eliminating severe false-positive predictions are achieved only after the phase-coherent temporal filtering plus spatial connectivity step. The justification rests solely on the untested assumption that 'genuine anatomy appears in contiguous blocks of phases, whereas transient artifacts appear sporadically,' with no quantitative verification on the evaluation data and no ablation isolating the filtering contribution. This makes it impossible to attribute the reported suppression of artifacts and retained accuracy (>95% of full-data) to the LoRA adaptation alone.
  2. [Abstract] Abstract: The reported numeric results lack accompanying information on total dataset size, cross-validation procedure, additional baselines beyond zero-shot SAM 3, or statistical testing. Without these, the strength of the claims (e.g., median Dice 0.968/0.910 and retention of 95% accuracy) cannot be assessed for robustness or generalizability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thoughtful and constructive comments, which highlight important aspects of clarity and substantiation in our work. We address each major comment below and commit to revisions that strengthen the manuscript without misrepresenting our contributions.

read point-by-point responses

  1. Referee: [Abstract] Abstract (inference stage): The headline performance metrics and the claim of eliminating severe false-positive predictions are achieved only after the phase-coherent temporal filtering plus spatial connectivity step. The justification rests solely on the untested assumption that 'genuine anatomy appears in contiguous blocks of phases, whereas transient artifacts appear sporadically,' with no quantitative verification on the evaluation data and no ablation isolating the filtering contribution. This makes it impossible to attribute the reported suppression of artifacts and retained accuracy (>95% of full-data) to the LoRA adaptation alone.

    Authors: We agree that the reported metrics reflect the full pipeline, including phase-coherent filtering. The filtering is motivated by the established physical properties of periodic respiratory motion in 4DCT, where true anatomical structures exhibit temporal contiguity across phases while artifacts do not. However, to enable clearer attribution, we will add an ablation study that isolates the filtering step (reporting Dice and HD95 with and without it) and include quantitative verification by computing phase-contiguity metrics on both ground-truth annotations and zero-shot predictions within the evaluation set. This will be presented in a new results subsection. revision: yes

  2. Referee: [Abstract] Abstract: The reported numeric results lack accompanying information on total dataset size, cross-validation procedure, additional baselines beyond zero-shot SAM 3, or statistical testing. Without these, the strength of the claims (e.g., median Dice 0.968/0.910 and retention of 95% accuracy) cannot be assessed for robustness or generalizability.

    Authors: We will revise the abstract to explicitly state the total dataset size (number of 4DCT volumes), the cross-validation procedure (patient-wise), and include a brief note on statistical testing (e.g., reporting inter-quartile ranges or Wilcoxon tests for key comparisons). The full manuscript already details the seven annotated volumes used for adaptation and the full-data baseline for the 95% retention claim; we will ensure these elements are summarized concisely in the abstract and add any feasible additional baselines (such as a standard U-Net) if space permits, or reference them in the main text. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical results independent of inputs

full rationale

The paper reports experimental Dice/Hausdorff metrics after LoRA fine-tuning on 7 volumes plus inference-time filtering. The filtering step invokes a domain assumption about phase-contiguous anatomy versus sporadic artifacts, but this is stated as prior knowledge rather than derived from any equation or fit within the paper. No equations, fitted parameters renamed as predictions, or self-citation chains appear in the provided text. Performance numbers are measured on held-out data and do not reduce to the method inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract supplies only high-level domain assumptions; no free parameters, invented entities, or additional axioms are stated.

axioms (1)
  • domain assumption Respiratory motion is continuous and periodic so that genuine anatomy appears in contiguous phase blocks while artifacts are sporadic.
    Invoked to justify the temporal filtering step described in the abstract.

pith-pipeline@v0.9.1-grok · 5776 in / 1342 out tokens · 40429 ms · 2026-06-27T04:30:55.419775+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

25 extracted references · 12 canonical work pages · 6 internal anchors

  1. [1]

    In: Seminars in radiation oncology

    Cardenas, C.E., Yang, J., Anderson, B.M., Court, L.E., Brock, K.B.: Advances in auto-segmentation. In: Seminars in radiation oncology. vol. 29, pp. 185–197. Elsevier (2019)

    2019

  2. [2]

    Carion, N., Gustafson, L., Hu, Y.T., Debnath, S., Hu, R., Suris, D., Ryali, C., Al- wala, K.V., Khedr, H., Huang, A., Lei, J., Ma, T., Guo, B., Kalla, A., Marks, M., Greer, J., Wang, M., Sun, P., Rädle, R., Afouras, T., Mavroudi, E., Xu, K., Wu, T.H., Zhou, Y., Momeni, L., Hazra, R., Ding, S., Vaze, S., Porcher, F., Li, F., Li, S., Kamath, A., Cheng, H.K....

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  3. [3]

    Nature machine intelligence5(3), 220–235 (2023)

    Ding, N., Qin, Y., Yang, G., Wei, F., Yang, Z., Su, Y., Hu, S., Chen, Y., Chan, C.M., Chen, W., et al.: Parameter-efficient fine-tuning of large-scale pre-trained language models. Nature machine intelligence5(3), 220–235 (2023)

    2023

  4. [4]

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale (2021), https://arxiv.org/abs/2010.11929

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  5. [5]

    https://doi.org/10.7937/3PPX-7S22, https://www.cancerimagingarchive.net/collection/ct-vs-pet-ventilation-imaging/

    Eslick, E.M., Kipritidis, J., Gradinscak, D., Stevens, M.J., Bai- ley, D.L., Harris, B., Booth, J.T., Keall, P.J.: Ct ventilation as a functional imaging modality for lung cancer radiotherapy (ct-vs- pet-ventilation-imaging) (2022). https://doi.org/10.7937/3PPX-7S22, https://www.cancerimagingarchive.net/collection/ct-vs-pet-ventilation-imaging/

    work page doi:10.7937/3ppx-7s22 2022

  6. [6]

    Hatamizadeh, A., Tang, Y., Nath, V., Yang, D., Myronenko, A., Landman, B., Roth, H., Xu, D.: Unetr: Transformers for 3d medical image segmentation (2021), https://arxiv.org/abs/2103.10504

    work page arXiv 2021

  7. [7]

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models (2021), https://arxiv.org/abs/2106.09685

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  8. [8]

    International Commission on Radiation Units and Measure- ments (1999)

    International Commission on Radiation Units and Measurements: ICRU Report 62: Prescribing, Recording and Reporting Photon Beam Therapy (Supplement to ICRU Report 50). International Commission on Radiation Units and Measure- ments (1999)

    1999

  9. [9]

    Nature methods18(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 methods18(2), 203–211 (2021)

    2021

  10. [10]

    Medical physics33(10), 3874–3900 (2006)

    Keall, P.J., Mageras, G.S., Balter, J.M., Emery, R.S., Forster, K.M., Jiang, S.B., Kapatoes, J.M., Low, D.A., Murphy, M.J., Murray, B.R., et al.: The management of respiratory motion in radiation oncology report of aapm task group 76 a. Medical physics33(10), 3874–3900 (2006)

    2006

  11. [11]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4015–4026 (2023)

    2023

  12. [12]

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

    2017

  13. [13]

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization (2019), https://arxiv.org/abs/1711.05101 10 C. Song

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  14. [14]

    Nature communications15(1), 654 (2024)

    Ma, J., He, Y., Li, F., Han, L., You, C., Wang, B.: Segment anything in medical images. Nature communications15(1), 654 (2024)

    2024

  15. [15]

    Milletari, F., Navab, N., Ahmadi, S.A.: V-net: Fully convolutional neural networks for volumetric medical image segmentation (2016), https://arxiv.org/abs/1606.04797

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  16. [16]

    SAM 2: Segment Anything in Images and Videos

    Ravi, N., Gabeur, V., Hu, Y.T., Hu, R., Ryali, C., Ma, T., Khedr, H., Rädle, R., Rolland, C., Gustafson, L., et al.: Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  17. [17]

    Robinson, J., Chuang, C.Y., Sra, S., Jegelka, S.: Contrastive learning with hard negative samples (2021), https://arxiv.org/abs/2010.04592

    work page arXiv 2021

  18. [18]

    In: International Conference on Medical image computing and computer-assisted intervention

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedi- cal image segmentation. In: International Conference on Medical image computing and computer-assisted intervention. pp. 234–241. Springer (2015)

    2015

  19. [19]

    International Journal of Radiation Oncology* Biology* Physics 63(1), 253–260 (2005)

    Underberg, R.W., Lagerwaard, F.J., Slotman, B.J., Cuijpers, J.P., Senan, S.: Use of maximum intensity projections (mip) for target volume generation in 4dct scans for lung cancer. International Journal of Radiation Oncology* Biology* Physics 63(1), 253–260 (2005)

    2005

  20. [20]

    Physics in Medicine & Biology48(1), 45–62 (2003)

    Vedam, S., Keall, P., Kini, V., Mostafavi, H., Shukla, H., Mohan, R.: Acquiring a four-dimensional computed tomography dataset using an external respiratory signal. Physics in Medicine & Biology48(1), 45–62 (2003)

    2003

  21. [21]

    Radiology: Artificial Intelligence 5(5), e230024 (2023)

    Wasserthal, J., Breit, H.C., Meyer, M.T., Pradella, M., Hinck, D., Sauter, A.W., Heye, T., Boll, D.T., Cyriac, J., Yang, S., et al.: Totalsegmentator: robust segmen- tation of 104 anatomic structures in ct images. Radiology: Artificial Intelligence 5(5), e230024 (2023)

    2023

  22. [22]

    Wu, J., Ji, W., Liu, Y., Fu, H., Xu, M., Xu, Y., Jin, Y.: Medical sam adapter: Adapting segment anythingmodel formedical image segmentation (2023), https://arxiv.org/abs/2304.12620

    work page arXiv 2023

  23. [23]

    International Journal of Radiation Oncology* Biology* Physics38(1), 197–206 (1997)

    Yan, D., Wong, J., Vicini, F., Michalski, J., Pan, C., Frazier, A., Horwitz, E., Mar- tinez, A.: Adaptive modification of treatment planning to minimize the deleterious effects of treatment setup errors. International Journal of Radiation Oncology* Biology* Physics38(1), 197–206 (1997)

    1997

  24. [24]

    Zhang, K., Liu, D.: Customized segment anything model for medical image seg- mentation (2023), https://arxiv.org/abs/2304.13785

    work page arXiv 2023

  25. [25]

    Zhu, J., Hamdi, A., Qi, Y., Jin, Y., Wu, J.: Medical sam 2: Segment medical images as video via segment anything model 2 (2024), https://arxiv.org/abs/2408.00874

    work page arXiv 2024