Implicit Shape-Prior for Few-Shot Assisted 3D Segmentation

Ana\"is Barateau; Anne-Sophie Boureau; Antoine Nordez; Blanche Texier; C\'edric H\'emon; Diana Mateus; Jean-Claude Nunes; J\'er\'emie Huet; Louise Piecuch; Mathilde Monvoisin

REVIEW 3 major objections 5 minor 24 references

Implicit Shape-Prior for Few-Shot Assisted 3D Segmentation

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read This paper claims that selecting annotation slices where the model's predictions err most—rather than spacing them evenly—yields more accurate multi-organ 3D segmentations from only a few manual slices, demonstrated on brain CT and sarcopen

desk verdict A solid, clearly-written incremental extension of auto-decoder shape priors to multi-organ few-shot segmentation, but the headline comparison does not isolate error-based slice selection from anatomy-aware slice placement. read the letter →

arxiv 2509.08580 v1 pith:FQBFNJZQ submitted 2025-09-10 cs.CV cs.LG

Mathilde Monvoisin , Louise Piecuch , Blanche Texier , Cédric Hémon , Anaïs Barateau , Jérémie Huet , Antoine Nordez , Anne-Sophie Boureau

show 2 more authors

Jean-Claude Nunes Diana Mateus

This is my paper · ORCID

classification cs.CVcs.LG

keywords Semi-automaticsegmentationImplicitshapepriorsNeuralauto-decoderSliceselectionOrgan-at-riskSarcopenia3DultrasoundFew-shot

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Manual 3D segmentation is a bottleneck in radiotherapy planning and sarcopenia diagnosis. This paper tries to show that a learned shape prior—a compact neural model of expected organ geometry—can reconstruct an entire 3D volume from only two or three expert-annotated slices, and that the right choice of slices matters more than slice count. Its proposed strategy is to start from one middle slice per organ and then add the slice where the model's prediction is farthest from the truth, rather than spacing slices evenly. The authors demonstrate the approach on brain CT organs at risk and on sarcopenic rectus femoris muscles in 3D ultrasound, reporting improved Dice, surface distance, and Hausdorff metrics when very few slices are available. If correct, the method would let clinicians cut annotation effort substantially while keeping or improving segmentation accuracy.

What carries the argument

The load-bearing object is the implicit shape prior implemented as an auto-decoder: one multilayer perceptron fθ that maps a 3D coordinate x plus a latent shape code z to per-class occupancy probabilities. Training optimizes both the network weights and one latent vector per training volume against full segmentation masks with a Dice plus cross-entropy loss; inference freezes the weights and optimizes only the latent code of a new volume against the annotated slices. The second mechanism is slice selection from error maps. In UC1, predictions from middle-slice initialization are subtracted from ground truth on training volumes, averaged into one error map, and the slice with highest mean err

What would settle it

Compare held-out test patients at equal slice budgets: run the method's training-set or adaptation-set slice selection against evenly spaced slices (and against random slices) over several cross-validation folds, measuring Dice and Hausdorff distance. If even or random spacing matches or beats error-based selection on average, or if per-patient ground-truth error maps do not rank the selected slices near the top, the central claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that an implicit shape prior learned from a small set of segmented volumes can turn a handful of manually annotated 2D slices into an accurate full 3D segmentation, and that the choice of which slices to annotate matters as much as the number. The authors extend a single-organ auto-decoder—a coordinate-based network that predicts per-voxel class probabilities from 3D coordinates and a per-patient latent vector—to multi-organ segmentation, then select slices by comparing initial predictions with ground truth and adding the slice with the highest error. On 20 registered brain CTs (8 training, 10 test), this error-guided strategy beats regular slice spacing at every

Load-bearing premise

The load-bearing premise is that a shape prior learned from a few segmented volumes can stretch to fit unseen and abnormal shapes, and that the places where the model makes mistakes on training or adaptation patients are the same places it will make mistakes on future patients.

Editorial extensions

If this is right

With two or three annotated slices, error-map-based slice selection improves average Dice, average surface distance, and maximum Hausdorff distance relative to evenly spaced slices on the brain CT use case; the largest gain is at two slices (Dice 0.70 vs 0.55).
On the sarcopenia use case, the method lowers Hausdorff distance and volumetric error with three to six slices, keeping Hausdorff around 15 mm where the baseline exceeds 20 mm.
Because the method constrains slice selection to the object of interest, it avoids wasted annotations on background slices, which the baseline sometimes selects.
The benefits concentrate in the low-slice regime; gains saturate as more slices are added, and in UC1 the selected slices can be very close together.
The multi-organ extension allows a single middle slice per organ to initialize the segmentation and then focuses extra annotations where errors are highest.

Reading between the lines

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

The error-map logic could be turned into an online per-patient policy that uses the model's current probability maps to choose the next slice; the paper only sketches uncertainty as a future direction but does not test it.
The success on sarcopenic muscles suggests the prior is not memorizing one morphology; a stronger test would be to measure how much accuracy degrades as the target population moves further from the training distribution, for example by grading sarcopenia severity clinically.
Because the method is image-modality-agnostic, the same mechanism should transfer to other sparse-annotation tasks such as vessel or tendon segmentation where insertion slices carry critical information; this is a direct corollary the paper does not claim.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Bottom line: this is a solid incremental method paper. The authors take Amiranashvili et al.'s implicit auto-decoder shape prior, generalize it from single-label to multi-organ, and add an error-map-driven slice-selection strategy. They evaluate on two real clinical problems—brain CT organs-at-risk and sarcopenic rectus femoris ultrasound—and the writing is clear. The multi-organ extension is natural but necessary; the slice-selection idea is simple and clinically sensible.

What the paper does well: it validates on genuinely hard data, including domain shift from healthy to sarcopenic muscles. The qualitative results match the quantitative trends. The authors are honest about the mixed DSC in UC2, and they explicitly note that the baseline sometimes samples background slices.

Where it falls short: the central claim—that choosing slices by highest prediction error beats regular spacing—is not established by the current comparisons. Both use cases confound error-based selection with anatomy-aware placement. In UC1 the proposed strategy always includes the middle slice of each organ; regular spacing across the whole volume can easily miss small organs or hit background. In UC2 the baseline samples the entire labelmap, while the proposed method constrains slices to the muscle and includes insertions and the mid-axial slice. The observed gains at low slice counts are consistent with the simpler explanation: guaranteed anatomical coverage. A control using evenly spaced slices within the organ's bounding box or normalized muscle length is missing. Without it, the incremental benefit of the error criterion itself is unknown.

There are smaller issues: no significance tests, boxplots show large overlap, and per-organ Table 1 shows the right-eye DSC is worse at two slices (0.24 vs 0.54) even though the aggregate 'better on every metric' claim holds. In UC2 it is not stated explicitly whether the three adaptation volumes are excluded from the inference test set; that should be clarified. Datasets and code are not public, which limits reproducibility.

Bottom line: the paper is worth a serious review. The method is plausible, the clinical motivation is real, and the multi-organ extension plus slice-selection framing is worth having. But the experiments need a control that separates 'knows where the anatomy is' from 'knows where the error is.' I would send it to peer review with that demand.

Referee Report

3 major / 5 minor

Summary. The paper extends the implicit shape-prior auto-decoder of Amiranashvili et al. [2] to multi-organ segmentation and proposes an interactive slice-selection strategy in which, after an initial set of manually annotated slices, the expert is guided to annotate slices where the model makes the largest prediction errors. The method is evaluated in two use cases: multi-organ segmentation of organs at risk in brain CT (UC1) and few-shot annotation of sarcopenic rectus femoris muscles in 3D ultrasound with a shape prior trained on non-sarcopenic subjects (UC2). The authors report that the proposed strategy outperforms baseline regular-spacing slice selection, especially at very low slice counts, and argue that it aligns better with clinical practice.

Significance. If confirmed, the contribution is practically relevant: it would reduce the number of manual slices needed for accurate 3D segmentation in radiotherapy planning and in the construction of muscle databases, while avoiding explicit correspondences or templates. The paper also addresses multi-organ extension of implicit shape priors, which is a useful step beyond prior single-structure work. Strengths include the use of two diverse medical datasets, a reproducible methodology that builds explicitly on prior work, and a clear description of the interactive workflow. However, the central claim—that error-based slice selection is superior to regular spacing—is not established by the current experimental design, because the comparison conflates error-based selection with anatomy-aware initialization and because no statistical significance tests are provided. The contributions are therefore promising but need a more controlled evaluation.

major comments (3)

[§2.4, §3.1, §3.2] The comparison is confounded: strategy 1 always includes the middle slice of each organ (UC1) or the insertions and mid-axial slice (UC2), while strategy 0 uses regular spacing across the entire volume, which can fall in background or miss small structures. As the authors note in §3.2, the baseline sometimes selects slices outside the actual segmentation area. The observed gains at low slice counts are therefore consistent with the simpler hypothesis that strategy 1 guarantees every annotated slice intersects the target anatomy, not that the highest-error criterion is beneficial. An ablation comparing error-based selection against regular spacing within the organ's bounding box (or along normalized muscle length) is necessary to attribute the improvement to the error-based mechanism. Without this control, the paper's central claim is not established.
[§3.2, Fig. 6] The claim that 'our method demonstrates superior performance on the other two metrics' is selective: DSC is lower for the proposed strategy at most slice counts (e.g., slices 3, 4, 5, 7, 8 and 9 per Fig. 6), while only HSD and volumetric error are better at some counts. No statistical significance tests are reported, and the boxplots show large overlap and variance across both strategies, especially for DSC and HSD. Given the small test set (the exact number of sarcopenic test volumes is not stated), the evidence that the proposed strategy is 'better' is not statistically supported. At minimum, paired tests or confidence intervals per slice count and metric are needed.
[Table 1, Sec 3.1] The aggregate claim that strategy 1 'outperforms the baseline on every metric when inferring with two or three slices' hides important per-organ regressions. For example, at 2 slices the right-eye DSC is 0.24 with strategy 1 versus 0.54 with strategy 0, and at 3 slices the right-eye DSC is 0.27 versus 0.33. Thus the aggregate improvement is driven mainly by the larger organs (spinal cord, brain). The paper should report per-metric, per-organ results consistently, and the discussion should acknowledge that the proposed selection is not uniformly beneficial for small or laterally located structures.

minor comments (5)

[General] There are several typos and grammatical issues: 'developped' (§3), 'startegy' (Fig. 7 and text), 'Superpostion' (Fig. 7 caption), 'c.f.' should be 'cf.' These do not affect the technical content but should be corrected.
[§2.4, UC1] The slice selection for UC1 is performed on the training set and then fixed for all test patients. Since patient registration is used, this is a reasonable choice, but the paper should clarify whether the same global slice indices were applied to all test volumes or whether the middle-slice computation is per-patient after registration. This detail is important for reproducibility.
[§2.4, UC2] The UC2 selection procedure involves several hand-designed choices (zone splitting into three, alternating between zones 1 and 3, minimum gap of 5 slices). It is unclear how sensitive the results are to these choices; a sensitivity analysis or at least a statement of the rationale would strengthen the paper.
[§3.1, Fig. 4] The boxplots show aggregate metrics (DSC, ASD, Hausdorff) but the unit of aggregation (voxel-level? volume-level? organ-level?) is not defined. Readers should know whether DSC is computed per organ and averaged over organs, or computed on the multiclass volume. The table caption for Table 1 says 'Average DSC' but the aggregation across organs is ambiguous.
[References] The paper relies heavily on [2] for architectural details; it would be helpful to state clearly which hyperparameters are inherited versus newly set. Also, the baseline experiments in UC2 are only compared against [2]; a comparison with a modern interactive segmentation method (e.g., SISeg [17]) is missing, although the authors acknowledge the difference in scope.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the slice-selection and inference pipeline is evaluated on held-out test volumes.

full rationale

The paper's derivation chain is self-contained with respect to its central claim. An implicit decoder is trained on full 3D labelmaps (Eq. 1) to learn a population shape prior; at inference, a new latent code is optimized against only the expert-annotated sparse slices (Sec. 2.3). The slice-selection strategy (Sec. 2.4) computes error maps by comparing the trained model's predictions with ground truth on training volumes (UC1) or three adaptation volumes (UC2), and then fixes the chosen slice indices before applying them to held-out test subjects. The quantitative comparison (Secs. 3.1, 3.2) therefore measures reconstruction quality on test volumes whose ground truth was not used to choose the slices or to fit any parameter. No equation, constant, or prediction is defined in terms of the method's own output. The paper builds on Amiranashvili et al. [2] as an external baseline and prior architecture; this is ordinary extension, not self-citation. The cited works by overlapping authors [13,14,20] support data acquisition and dataset-composition conventions, but none carries a load-bearing uniqueness argument or substitutes for an independent derivation. The observation that the baseline's regular spacing can fall in background (Sec. 3.2) is a potential comparison confound and a validity concern, but it is not circularity under the definitions used here. Accordingly, no circular step is identified.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on design choices for slice selection (zones, alternation, initialization), hyperparameters inherited from prior work, and domain assumptions about shape-prior transfer. No new physical entities or constants are introduced.

free parameters (4)

Slice-selection hyperparameters for UC2 (zone splitting, alternation, minimum gap)
The muscle is split into three zones, slices alternate between zones 1 and 3, and a constraint of at least 5 slices from previous selections is imposed. These rules are chosen by hand, not justified by an ablation.
Initial slice rule (middle slice per organ in UC1; insertions and mid-axial in UC2)
The minimal starting subset is defined by anatomical heuristics rather than learned or optimized, and affects all subsequent selections.
Latent code dimension 128*Nclass
The per-volume shape latent dimension is inherited from [2] and scaled by the number of classes; it is a hand-chosen architectural choice.
Inference epochs (300 for UC1, 1500 for UC2)
The number of optimization steps for the latent code at inference is set by observing convergence; it affects reconstruction quality and the error maps used for slice selection.

assumptions (4)

domain assumption The learned implicit shape prior can represent the anatomical variability of the target population, including abnormal sarcopenic shapes, through latent code optimization.
Used in Sec 2.4 for UC2: a prior trained on normal subjects must reconstruct abnormal shapes. If the prior's latent manifold cannot express these shapes, few-shot inference fails regardless of slice selection.
domain assumption Error maps computed on the training set (UC1) or on 3 adaptation volumes (UC2) generalize to unseen patients.
In Sec 3.1 and 3.2, slice indices are fixed once and applied to all test patients. If the training error map is not representative of test anatomy, the selected slices may be suboptimal.
domain assumption Rigid registration (UC1) and muscle-length normalization (UC2) align anatomies sufficiently for a coordinate-based implicit model to share a population prior.
The auto-decoder relies on coordinates in a common reference space; if registration or normalization is inaccurate, the population prior is misaligned across subjects.
domain assumption The auto-decoder inference procedure, optimizing a new latent code with sparse slice losses, converges to a shape on the same manifold as training.
Sections 2.2-2.3 assume that minimizing the slice loss with respect to z_new finds an accurate full-volume shape; this is the standard auto-decoder assumption and is not independently proven here.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Implicit Shape-Prior for Few-Shot Assisted 3D Segmentation." pith.science (2026). https://pith.science/paper/FQBFNJZQ

@misc{pith2026250908580,
  author       = {Pith},
  title        = {Pith review of: Implicit Shape-Prior for Few-Shot Assisted 3D Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FQBFNJZQ}},
  note         = {Machine review of arXiv:2509.08580}
}

read the original abstract

The objective of this paper is to significantly reduce the manual workload required from medical professionals in complex 3D segmentation tasks that cannot be yet fully automated. For instance, in radiotherapy planning, organs at risk must be accurately identified in computed tomography (CT) or magnetic resonance imaging (MRI) scans to ensure they are spared from harmful radiation. Similarly, diagnosing age-related degenerative diseases such as sarcopenia, which involve progressive muscle volume loss and strength, is commonly based on muscular mass measurements often obtained from manual segmentation of medical volumes. To alleviate the manual-segmentation burden, this paper introduces an implicit shape prior to segment volumes from sparse slice manual annotations generalized to the multi-organ case, along with a simple framework for automatically selecting the most informative slices to guide and minimize the next interactions. The experimental validation shows the method's effectiveness on two medical use cases: assisted segmentation in the context of at risks organs for brain cancer patients, and acceleration of the creation of a new database with unseen muscle shapes for patients with sarcopenia.

Figures

Figures reproduced from arXiv: 2509.08580 by the authors.

**Figure 1.** Implicit shape decoder. The training phase (left) provides the full segmented volume as target to learn the latent shape vector zi. The inference phase (right) takes only some slices as input to predict a full segmentation with the previously learned zi. 2.1 Implicit Prior Shape Learning Our implicit model consist of a single Multilayer Perceptron (MLP) used as a pervoxel classifier fθ, to obtain a binary or multi-… view at source ↗

**Figure 2.** Overview of error map generation and slice selection for UC1 Use case 2 (UC2) : Few-Shot Annotation of a New Database with Domain Shift This use case targets the creation of a new database involving ‘abnormal’ muscle shapes (e.g., sarcopenia), adapting the shape prior from a model trained on non-sarcopenic subjects. As shown in [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Overview of combined error map generation and slice selection for UC2 (top) and detailed metric calculation and slice selection (bottom). 3 Experiments Implementation Details Most of the model parameters are kept as in [2] including the MLP architecture: the learning rate was set to 1.0e −4 and to 1.0e −3 for the AD. The MLP has 8 layers, the output of the 4th layer is concatenated with the 3D coordinates, and the d… view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Comparison across different slice numbers on the test set [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: Slices and predictions used for each amount of slices. Cyan is the brain, dark blue is the left eye, yellow is the right eye and green is the spinal cord. Discussion The proposed interactive method is more robust and achieves better scores than the baseline across all …

**Figure 6.** Figure 6: Comparison across different slice numbers on the test set [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

**Figure 7.** Figure 7: Comparison of qualitative results on different number of slices of Amiranashvili strategy (bottom) with ours (top) on UC2. Supperpostion of prediction in red, with the GT slices inputed to the model in dark green and the GT shape in lightgreen. Discussion In UC2, which…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 4 linked inside Pith

[2]

Amiranashvili, T., Lüdke, D., Li, H.B., Zachow, S., Menze, B.H.: Learning con- tinuous shape priors from sparse data with neural implicit functions. Med. Img. Analysis94(2024)

2024
[1]

Elhabian, S.: Weakly supervised bayesian shape modeling from unsegmented medical images

Adams, J., Iyer, K., Y. Elhabian, S.: Weakly supervised bayesian shape modeling from unsegmented medical images. In: Int. Workshop on Shape in Med. Imaging. Springer (2024)

2024
[3]

Bhalodia, R., Elhabian, S., Adams, J., Tao, W., Kavan, L., Whitaker, R.: Deepssm: A blueprint for image-to-shape deep learning models. Med. Img. Analysis91(2024)

2024
[4]

Scientific reports13(1) (2023) Implicit Shape-Prior for Few-Shot Assisted 3D Segmentation 15

Bongratz, F., Rickmann, A.M., Wachinger, C.: Abdominal organ segmentation via deep diffeomorphic mesh deformations. Scientific reports13(1) (2023) Implicit Shape-Prior for Few-Shot Assisted 3D Segmentation 15

2023
[5]

In: IEEE Engineering in Medicine and Biology Society (EMBC)

Cerqueira, M.A., Sprenger, F., Teixeira, B.C., Falcão, A.X.: Interactive image se- lection and training for brain tumor segmentation network. In: IEEE Engineering in Medicine and Biology Society (EMBC). IEEE (2024)

2024
[6]

Chang, A., Tao, X., Yang, X., Huang, Y., Zhou, X., Zeng, J., Huang, R., Ni, D.: Pe-med: Prompt enhancement for interactive medical image segmentation. In: Int. workshop on machine learning in medical imaging. Springer (2023)

2023
[7]

arXiv preprint arXiv:2308.16184 (2023)

Cheng, J., Ye, J., Deng, Z., Chen, J., Li, T., Wang, H., Su, Y., Huang, Z., Chen, J., Jiang, L., et al.: Sam-med2d. arXiv preprint arXiv:2308.16184 (2023)

arXiv 2023
[8]

Physics in Medicine & Biology70(7) (2025)

Cubero, L., Hémon, C., Barateau, A., Castelli, J., de Crevoisier, R., Acosta, O., Pascau, J.: Deep learning-based segmentation of head and neck organs at risk on cbct images with dosimetric assessment for radiotherapy. Physics in Medicine & Biology70(7) (2025)

2025

Show all 24 references

[9]

De Paolis, G.R., Lenis, D., Novotny, J., Wimmer, M., Berg, A., Neubauer, T., Matthias Winter, P., Major, D., Muthusami, A., Schröcker, G., et al.: Fast medi- cal shape reconstruction via meta-learned implicit neural representations. In: Int. Workshop on Shape in Med. Imaging. ...

2024
[10]

In: MICCAI Workshop on Data Augmentation, Labelling, and Imperfections

Diaz-Pinto, A., Mehta, P., Alle, S., Asad, M., Brown, R., Nath, V., Ihsani, A., Antonelli, M., Palkovics, D., Pinter, C., et al.: Deepedit: Deep editable learning for interactive segmentation of 3d medical images. In: MICCAI Workshop on Data Augmentation, Labelling, and Imperf...

2022
[11]

In: IEEE/CVF Workshop on Applications of Computer Vision (WACV) (2022)

Hatamizadeh, A., Tang, Y., Nath, V., Yang, D., Myronenko, A., Landman, B., Roth, H.R., Xu, D.: Unetr: Transformers for 3d medical image segmentation. In: IEEE/CVF Workshop on Applications of Computer Vision (WACV) (2022)

2022
[12]

Huang, Y., Yang, X., Liu, L., Zhou, H., Chang, A., Zhou, X., Chen, R., Yu, J., Chen, J., Chen, C., et al.: Segment anything model for medical images? Med. Img. Analysis92(2024)

2024
[13]

Journal of Biomechanics162(2024)

Huet, J., Boureau, A.S., Sarcher, A., Cornu, C., Nordez, A.: Validation of a scan- ning technique with minimal compression for measuring muscle volume with free- hand 3d ultrasound. Journal of Biomechanics162(2024)

2024
[14]

Huet, J., Nordez, A., Sarcher, A., Mathieu, M., Cornu, C., Boureau, A.: Concor- dance of freehand 3D-ultrasound muscle measurements with sarcopenia parameters in a geriatric rehabilitation ward. In: J. of Cachexia, Sarcopenia and Muscle (2024)

2024
[15]

In: IEEE Int

Jouvencel, M., Kéchichian, R., Digne, J., Valette, S.: Sconet: Convolutional occu- pancy networks for multi-organ segmentation. In: IEEE Int. Symp. on Biomedical Imaging (ISBI) (2025)

2025
[16]

In: IEEE/CVF Int

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: IEEE/CVF Int. Conf. on Computer Vision (ICCV). pp. 4015–4026 (2023)

2023
[17]

preprint arXiv:2411.19447 (2024)

Li, Z., Zhao, K., Wang, Y., Wang, S.: Adaptive interactive segmentation for mul- timodal medical imaging via selection engine. preprint arXiv:2411.19447 (2024)

2024 arXiv
[18]

Plos one 19(10) (2024)

Lin, Z., Dall’Ara, E., Guo, L.: A novel mean shape based post-processing method for enhancing deep learning lower-limb muscle segmentation accuracy. Plos one 19(10) (2024)

2024
[19]

Computer Methods and Programs in Biomedicine245(2024)

Mikhailov, I., Chauveau, B., Bourdel, N., Bartoli, A.: A deep learning-based inter- active medical image segmentation framework with sequential memory. Computer Methods and Programs in Biomedicine245(2024)

2024
[20]

In: IEEE Int

Piecuch, L., Huet, J., Frouin, A., Nordez, A., Boureau, A.S., Mateus, D.: Unsuper- vised anomaly detection on implicit shape representations for sarcopenia detection. In: IEEE Int. Symp. on Biomedical Imaging (ISBI) (2025) 16 M. Monvoisin, L. Piecuch et al

2025
[21]

Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedi- cal image segmentation. In: Int. Conf. on Medical Image Computing and Computer Assisted Interventions (MICCAI). pp. 234–241. Springer (2015)

2015
[22]

arXiv preprint arXiv:2312.06072 (2023)

Tian, M., Chen, X., Gao, Y.: A dynamic interactive learning framework for auto- mated 3d medical image segmentation. arXiv preprint arXiv:2312.06072 (2023)

2023 arXiv
[23]

Ukey, J., Kataria, T., Elhabian, S.Y.: Massm: An end-to-end deep learning frame- work for multi anatomy statistical shape modeling directly from images. In: Int. Workshop on Shape in Med. Imaging. pp. 149–163. Springer (2024)

2024
[24]

preprint arXiv:2408.01620 (2024)

Zhu, J., Wu, J.: Meduhip: Towards human-in-the-loop medical segmentation. preprint arXiv:2408.01620 (2024)

2024 arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [2]

Amiranashvili, T., Lüdke, D., Li, H.B., Zachow, S., Menze, B.H.: Learning con- tinuous shape priors from sparse data with neural implicit functions. Med. Img. Analysis94(2024)

2024

[2] [1]

Elhabian, S.: Weakly supervised bayesian shape modeling from unsegmented medical images

Adams, J., Iyer, K., Y. Elhabian, S.: Weakly supervised bayesian shape modeling from unsegmented medical images. In: Int. Workshop on Shape in Med. Imaging. Springer (2024)

2024

[3] [3]

Bhalodia, R., Elhabian, S., Adams, J., Tao, W., Kavan, L., Whitaker, R.: Deepssm: A blueprint for image-to-shape deep learning models. Med. Img. Analysis91(2024)

2024

[4] [4]

Scientific reports13(1) (2023) Implicit Shape-Prior for Few-Shot Assisted 3D Segmentation 15

Bongratz, F., Rickmann, A.M., Wachinger, C.: Abdominal organ segmentation via deep diffeomorphic mesh deformations. Scientific reports13(1) (2023) Implicit Shape-Prior for Few-Shot Assisted 3D Segmentation 15

2023

[5] [5]

In: IEEE Engineering in Medicine and Biology Society (EMBC)

Cerqueira, M.A., Sprenger, F., Teixeira, B.C., Falcão, A.X.: Interactive image se- lection and training for brain tumor segmentation network. In: IEEE Engineering in Medicine and Biology Society (EMBC). IEEE (2024)

2024

[6] [6]

Chang, A., Tao, X., Yang, X., Huang, Y., Zhou, X., Zeng, J., Huang, R., Ni, D.: Pe-med: Prompt enhancement for interactive medical image segmentation. In: Int. workshop on machine learning in medical imaging. Springer (2023)

2023

[7] [7]

arXiv preprint arXiv:2308.16184 (2023)

Cheng, J., Ye, J., Deng, Z., Chen, J., Li, T., Wang, H., Su, Y., Huang, Z., Chen, J., Jiang, L., et al.: Sam-med2d. arXiv preprint arXiv:2308.16184 (2023)

arXiv 2023

[8] [8]

Physics in Medicine & Biology70(7) (2025)

Cubero, L., Hémon, C., Barateau, A., Castelli, J., de Crevoisier, R., Acosta, O., Pascau, J.: Deep learning-based segmentation of head and neck organs at risk on cbct images with dosimetric assessment for radiotherapy. Physics in Medicine & Biology70(7) (2025)

2025