Correction-aware interactive 3D tumor segmentation with sparse and revisable prompts

May combine two claims.

MSD-Colon contains 126 abdominal CT volumes with colon tumor annotations, and KiTS21 contains 300 CT volumes with kidney tumor annotations.

“MSD-Colon contains 126 abdominal CT volumes with colon tumor annotations, and KiTS21 contains 300 CT volumes with kidney tumor annotations.”

• cites [2] Antonelli, M., Reinke, A., Bakas, S., Farahani, K., Kopp-Schneider, A., Landman, B.A., Litjens, G., Menze, B., Ronneberger, O., Summers, R.M., et al.: The medical segmentation decathlon. Nat. Commun. 13, 4128 (2022). https://doi.org/10.1038/s41467-022-30695-9 sourcePartly supported

  Table 2 (Methods, The Decathlon mission; page unavailable) lists the Colon task as 190 CT volumes targeting colon cancer primaries, split into 126 training and 64 test volumes. Thus 126 is not the total stated in the claim. The paper’s dataset summary does not address KiTS21 or its claimed 300 kidney-tumor CT volumes. (LLM, conf=0.90)

  Experimental — accuracy not yet measured
• cites [8] Heller, N., Isensee, F., Trofimova, D., Tejpaul, R., Zhao, Z., Chen, H., Wang, L., Golts, A., Khapun, D., Shats, D., et al.: The KiTS21 challenge: automatic segmentation of kidneys, renal tumors, and renal cysts in corticomedullary-phase CT. arXiv preprint arXiv:2307.01984, (2023). https://doi.org/10.48550/arXiv.2307.01984 sourcePartly supported

  Section 2.1.1 (page unavailable) identifies 300 CT cases for the KiTS21 training set, supporting the kidney-tumor portion of the claim. Section 2.1.2 (page unavailable) describes a separate 100-case test set, so 300 is not the full KiTS21 dataset. The paper does not address the claimed 126 MSD-Colon volumes or their annotations. (LLM, conf=0.90)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

3DSAM-Adapter and SAM-Med3D adapt SAM-style prompting to 3D medical images, while ProMISe is a prompt-driven 3D medical segmentation method based on pretrained image foundation models.

“3DSAM-Adapter [5] and SAM-Med3D [29] adapt SAM-style prompting to 3D medical images, while ProMISe [17] is a prompt-driven 3D medical segmenta-tion method based on pretrained image foundation models.”

• cites [5] Gong, S., Zhong, Y., Ma, W., Li, J., Wang, Z., Zhang, J., Heng, P.-A., Dou, Q.: 3DSAM-adapter: holistic adaptation of SAM from 2D to 3D for promptable tumor segmentation. Med. Image Anal. 98, 103324 (2024). https://doi.org/10.1016/j.media.2024.103324 sourceSupported
• cites [29] Wang, H., Guo, S., Ye, J., Deng, Z., Cheng, J., Li, T., Chen, J., Su, Y., Huang, Z., Shen, Y., Fu, B., Zhang, S., He, J., Qiao, Y.: SAM-Med3D: Towards general-purpose segmentation models for volumetric medical images. arXiv preprint arXiv:2310.15161, (2023). https://doi.org/10.48550/arXiv.2310.15161 sourceSupported
• cites [17] Li, H., Liu, H., Hu, D., Wang, J., Oguz, I.: Promise: Prompt-driven 3D medical image segmentation using pretrained image foundation models. In 2024 IEEE 21st International Symposium on Biomedical Imaging (ISBI), pages 1–5. IEEE, (2024b). https://doi.org/10.1109/ISBI56570.2024.10635207 sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

nnU-Net is a widely used reference baseline for medical image segmentation because of its self-configuring training and preprocessing pipeline.

“nnU-Net is a widely used reference baseline for medical image segmentation because of its self-configuring training and preprocessing pipeline.”

• cites [9] Isensee, F., Jaeger, P.F., Kohl, S.A.A., Petersen, J., Maier-Hein, K.H.: nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat. Methods 18(2), 203–211 (2021). https://doi.org/10.1038/s41592-020-01008-z sourcePartly supported

  The Discussion (pp. 10–12) describes nnU-Net as a standardized baseline for comparing methods, and the Introduction (p. 2) explains how it automatically configures its pipeline. These passages support its suitability as a reference baseline, but they do not establish that it was already widely used as one. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

May combine two claims.

To make the comparison with PRISM controlled, we use the released PRISM data splits without resampling the train/test partition, resulting in 26 test cases for MSD-Colon and 61 test cases for KiTS21.

“To make the comparison with PRISM [16] controlled, we use the released PRISM data splits without resampling the train/test partition, resulting in 26 test cases for MSD-Colon and 61 test cases for KiTS21.”

• cites [16] Li, H., Liu, H., Hu, D., Wang, J., Oguz, I.: PRISM: a promptable and robust interactive segmentation model with visual prompts. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2024, volume 15003 of Lecture Notes in Computer Science, pages 389–399. Springer Nature Switzerland, (2024a). https://doi.org/10.1007/978-3-031-72384-1_37 sourcePartly supported

  The Experiments > Datasets passage says PRISM used colon and kidney datasets with 126 and 300 cases, respectively, and adopted a 70/10/20 training/validation/testing split from a prior study. It does not give the exact test counts of 26 and 61 or state that those partitions were released; the citing authors’ decision not to resample is their own procedure. (LLM, conf=0.82)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

3D UX-Net and Swin-UNETR represent recent high-capacity 3D architectures based on convolutional and transformer-style volumetric modeling.

“3D UX-Net and Swin-UNETR represent recent high-capacity 3D architectures based on convolutional and transformer-style volumetric modeling.”

• cites [13] Lee, H.H., Bao, S., Huo, Y., Landman, B.A.: 3D UX-Net: a large kernel volumetric convnet modernizing hierarchical transformer for medical image segmentation. In The Eleventh International Conference on Learning Representations, (2023). https://openreview.net/forum?id=wsZsjOSytRA sourceNot checked yet

  Checked before — its earlier verdict is held back until a new check runs.

• cites [27] Tang, Y., Yang, D., Li, W., Roth, H.R., Landman, B., Xu, D., Nath, V., Hatamizadeh, A.: Self-supervised pre-training of swin transformers for 3D medical image analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20730–20740, (2022). https://doi.org/10.1109/CVPR52688.2022.02007 sourcePartly supported

  Section 3, “Swin UNETR” (page not provided), describes a Swin Transformer encoder using 3D patches and a CNN-based decoder; the Models table lists 61.98 million parameters, supporting the description of Swin UNETR as a high-capacity volumetric architecture. The copy does not identify or describe 3D UX-Net as the convolutional counterpart (searched: “3D UX-Net” and “convolutional architectures high capacity 3D UX-Net”). (LLM, conf=0.88)

  Experimental — accuracy not yet measured

May combine two claims.

PRISM combines clicks, 3D bounding boxes, and scribbles with an iterative refinement head, making it a closely matched baseline for interactive 3D tumor segmentation.

“PRISM [16] combines clicks, 3D bounding boxes, and scribbles with an iterative refinement head, making it a closely matched baseline for interactive 3D tumor segmentation.”

• cites [16] Li, H., Liu, H., Hu, D., Wang, J., Oguz, I.: PRISM: a promptable and robust interactive segmentation model with visual prompts. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2024, volume 15003 of Lecture Notes in Computer Science, pages 389–399. Springer Nature Switzerland, (2024a). https://doi.org/10.1007/978-3-031-72384-1_37 sourceSupported

May combine two claims.

On MSD-Colon, the difference between the proposed method and PRISM appears from the early rounds and persists to the final prediction.

“On MSD-Colon, the difference appears from the early rounds and persists to the final prediction.”

May combine two claims.

Qualitative examples in Fig. 5 show that PRISM can leave residual boundary errors or produce small changes in the edited region across later rounds, while the proposed method follows the reference contour more closely and reaches a stable mask with fewer visible residual errors.

“PRISM can leave residual boundary errors or produce small changes in the edited region across later rounds.The proposed method follows the reference contour more closely in these examples and reaches a stable mask with fewer visible residual errors.”

May combine two claims.

Table 2 shows that the proposed method obtains higher Dice than PRISM at every tested prompt budget on both MSD-Colon and KiTS21.

“Table 2 shows that the proposed method obtains higher Dice than PRISM at every tested prompt budget on both datasets.”

May combine two claims.

Both PRISM and the proposed method improve as additional prompts are provided, but the proposed method remains above PRISM throughout the 11-round refinement sequence.

“Both PRISM and the proposed method improve as additional prompts are provided, but the proposed method remains above PRISM throughout the refinement sequence.”

May combine two claims.

The proposed method achieves higher segmentation accuracy across prompt settings, with larger differences when scribbles are limited to sparse informative slices and more consistent refinement when prompt labels are revised.

“The proposed method achieves higher segmentation accuracy across prompt settings, with larger differences when scribbles are limited to sparse informative slices and more consistent refinement when prompt labels are revised.”

May combine two claims.

Under the standard multi-prompt protocol, NSD changes little: 99.96% on MSD-Colon for both methods and 99.80% to 99.81% on KiTS21.

“NSD is already close to saturation under the standard pro-tocol and changes little: 99.96% on MSD-Colon for both methods and 99.80% to 99.81% on KiTS21.”

May combine two claims.

On KiTS21, the two curves are closer because the final Dice is already high for both methods, but the proposed method remains consistently higher.

“On KiTS21, the two curves are closer because the final Dice is already high for both methods, but the proposed method remains consistently higher.”

May combine two claims.

Under the standard multi-prompt protocol, the proposed method increases Dice from 93.79% to 95.43% on MSD-Colon and from 96.58% to 97.19% on KiTS21 kidney tumor.

“Under the same multi-prompt setting, the proposed method increases Dice from 93.79% to 95.43% on MSD-Colon and from 96.58% to 97.19% on KiTS21 kidney tumor.”

• Also stated in: § 7 Conclusion · plain

  “On MSD-Colon and KiTS21 kidney tumor, the framework improves Dice over PRISM under the standard multi-prompt protocol”

May combine two claims.

NSD changes less systematically across budgets because it is already high for both methods, especially on KiTS21.

“NSD changes less systematically across budgets because it is already high for both methods, especially on KiTS21.”

May combine two claims.

On MSD-Colon, most per-volume points lie above the diagonal with one, three, and five edited slices, with the clearest separation at the three- and five-slice budgets; under the dense-scribble protocol, the gap is smaller but the points remain largely on or above the diagonal.

“On MSD-Colon, most points lie above the diagonal with one, three, and five edited slices, with the clearest separation at the three-and five-slice budgets.Under the dense-scribble protocol, the gap is smaller but the points remain largely on or above the diagonal.”

May combine two claims.

The qualitative example shows a revision sequence in which this representation yields a smoother refinement trajectory, while the synthetic revision table shows higher final Dice under several levels of cross-round relabeling.

“The qualitative example shows a revision sequence in which this representation yields a smoother refinement trajectory, while the synthetic revision table shows higher final Dice under several levels of cross-round relabeling.”

May combine two claims.

The difference in final Dice is larger on MSD-Colon and becomes smaller as more prompt voxels are revised.

“The difference is larger on MSD-Colon and becomes smaller as more prompt voxels are revised, indicating that heavier synthetic revision creates a more challenging prompt-history configuration.”

May combine two claims.

On KiTS21, the points are more tightly clustered near the diagonal because both methods already obtain high Dice; even in this saturated regime, the sparse-budget panels show an upward shift for the proposed method, while the dense protocol shows only a modest difference.

“On KiTS21, the points are more tightly clustered near the diagonal because both methods already obtain high Dice.Even in this saturated regime, the sparse-budget panels show an upward shift for the proposed method, while the dense protocol shows only a modest difference.”

May combine two claims.

The effect is smaller on KiTS21, where both methods already operate in a high-Dice regime, but the same direction is observed across prompt budgets.

“The effect is smaller on KiTS21, where both methods already operate in a high-Dice regime, but the same direction is observed across prompt budgets.”

May combine two claims.

With only one edited slice, both methods operate under a severe correction budget, and the difference is correspondingly smaller.

“With only one edited slice, both methods operate under a severe correction budget, and the difference is correspondingly smaller.”

May combine two claims.

The proposed method more closely follows the reference contour over the refinement sequence, while PRISM retains residual boundary errors or changes near the lesion boundary across later rounds.

“The proposed examples show how the iterative curves in Fig. 4 appear visually: the proposed method more closely follows the reference contour over the refinement sequence, while PRISM retains residual boundary errors or changes near the lesion boundary across later rounds”

May combine two claims.

Figure 6 shows that with three or five edited slices, the proposed method separates from PRISM as refinement proceeds; with one edited slice, the curves are closer in the early rounds, but the final Dice remains higher for the proposed method.

“With three or five edited slices, the proposed method separates from PRISM as refinement proceeds.With one edited slice, the curves are closer in the early rounds, but the final Dice remains higher for the proposed method.”

May combine two claims.

In a revision sequence under sparse scribbles, PRISM exhibits a non-monotone Dice trajectory, whereas the proposed method changes more smoothly and ends with higher Dice.

“When prompt labels are revised across rounds, PRISM exhibits a non-monotone Dice trajectory, whereas the proposed method changes more smoothly and ends with higher Dice in this sequence.”

May combine two claims.

On KiTS21, the corresponding differences are smaller because the baseline is already strong, but the same trend is preserved.

“On KiTS21, the corresponding differences are smaller because the baseline is already strong, but the same trend is preserved.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The revision analysis is synthetic and is used to examine the prompt-memory representation, not to claim that the model is insensitive to arbitrary wrong prompts.

“The revision analysis is synthetic and is used to examine the prompt-memory representation, not to claim that the model is insensitive to arbitrary wrong prompts.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Randomizing the orientation prevents the model from depending on a single anatomical view: Compact and elongated tumors may be easier to correct from different planes.

“Randomizing the orientation prevents the model from depending on a single anatomical view: Compact and elongated tumors may be easier to correct from different planes.”

• AssumptionRests on an assumption: Tumor segmentation errors are often local, heterogeneous, and not well described by a single global object prior.

May combine two claims.

Common two-map memories accumulate foreground and background prompts separately; when a voxel is later relabeled, this creates an ambiguous input in which the same location can be active in both maps.

“Common two-map memories accumulate foreground and background prompts separately; when a voxel is later relabeled, this creates a ambiguous input in which the same location can be active in both maps.”

• AssumptionRests on an assumption: In a multi-round interaction, the prompt history can be revised.
• AssumptionRests on an assumption: A voxel marked as foreground in an early round may later be marked as background after the user sees an over-segmentation, or the reverse may occur for a missed tumor region.

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Dense-scribble protocols can tightly constrain a 3D mask, but they assume more user input than is available in many annotation-oriented editing sessions.

“Dense-scribble protocols can tightly constrain a 3D mask, but they assume more user input than is available in many annotation-oriented editing sessions.”

• AssumptionRests on an assumption: In practical correction, prompts are often sparse because only a few slices contain the most informative errors.
• AssumptionRests on an assumption: In a practical 3D editing session, the user is more likely to correct selected slices where the model is visibly wrong than to annotate the entire volume densely.

May combine two claims.

Under dense multi-prompt interaction, the main measurable difference between the proposed method and PRISM is in volumetric overlap; the surface-tolerance metric is already near ceiling for both interactive methods.

“Thus, under dense multi-prompt interaction, the main measurable difference is in volumetric overlap; the surface-tolerance metric is already near ceiling for both interactive methods.”

Probably not a substantial claim: may not pass S3 (load-bearing).

These methods provide a reference for how far fully automatic segmentation can go without user correction.

“These methods provide a reference for how far fully automatic segmentation can go without user correction.”

Classified as probably not a substantial claim.

These baselines are relevant because they test whether general or medical promptable models can handle challenging 3D tumor segmentation under point-based interaction.

“These baselines are relevant because they test whether general or medical promptable models can handle challenging 3D tumor segmentation under point-based interaction.”

Probably not a substantial claim: may not pass S3 (load-bearing).

For much longer editing sessions, C t could become saturated if many locations are repeatedly revised.

“For much longer editing sessions, however, C t could become saturated if many locations are repeatedly revised.”

Probably not a substantial claim: may not pass S2 (checkable).

This motivates interactive refinement rather than replacing automatic segmentation entirely.

“This motivates interactive refinement rather than replacing automatic segmentation entirely: The initial prediction can be useful, but the user needs an efficient way to correct the cases where it fails.”

Classified as probably not a substantial claim.

This keeps the experimental comparison focused on the interaction modeling changes rather than on preprocessing or data-split differences.

“This keeps the experimental comparison focused on the interaction modeling changes rather than on preprocessing or data-split differences.”

May combine two claims.

In this setting, the observed differences reflect the value of sparse residual-error prompting and revision-aware prompt-history modeling.

“In this setting, the observed differences reflect the value of sparse residual-error prompting and revision-aware prompt-history modeling.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The sparse-slice prompt distribution and revision-aware memory are evaluated as a joint framework; isolating their individual contributions would require additional retraining and ablation experiments.

“Third, the sparse-slice prompt distribution and revision-aware memory are evaluated as a joint framework.Isolating their individual contributions would require additional retraining and ablation experiments.”

May combine two claims.

Dense scribbles strongly constrain the mask and can reduce the need for learned correction behavior; sparse scribbles leave larger parts of the volume unconstrained and require the model to propagate limited user input through the 3D prediction.

“Dense scribbles strongly constrain the mask and can reduce the need for learned correction behavior.Sparse scribbles, in contrast, leave larger parts of the volume unconstrained and require the model to propagate limited user input through the 3D prediction.”

May combine two claims.

Ranking residual-error slices and training with sparse informative scribbles expose the model to the sparse regime.

“Ranking residual-error slices and training with sparse informative scribbles expose the model to this regime.”

May combine two claims.

Prompts are simulated from reference masks; this does not measure real annotation time, prompt placement variability, or user decision-making.

“Second, prompts are simulated from reference masks.This is standard for interactive segmentation benchmarking, but it does not measure real annotation time, prompt placement variability, or user decision-making.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The prompt-revision protocol is synthetic; it illustrates a refinement-input failure mode, but it is not a complete model of routine annotation-oriented editing.

“Fourth, the prompt-revision protocol is synthetic; it illustrates a refinement-input failure mode, but it is not a complete model of routine annotation-oriented editing.”

May combine two claims.

NSD is close to saturation in several settings, particularly for KiTS21, while Dice still changes with the prompt budget; this indicates that the observed improvements are mainly reflected in volumetric overlap and residual tumor volume rather than in large changes to the surface-tolerance metric.

“In several settings, NSD is close to saturation, particularly for KiTS21, while Dice still changes with the prompt budget.This indicates that the observed improvements are mainly reflected in volumetric overlap and residual tumor volume rather than in large changes to the surface-tolerance metric.”

May combine two claims. The quote only partly states this claim.

The improvement of the proposed method over PRISM is consistent but modest under the standard dense multi-prompt protocol, and increases when scribbles are restricted to a small number of informative slices, especially on MSD-Colon.

“When scribbles are restricted to a small number of informative slices, the difference increases, especially on MSD-Colon.”

Classified as probably not a substantial claim.

This choice means that the sparse-scribble results attribute performance to the complete correction-aware formulation rather than isolating a single component.

“This choice reflects the intended deployment regime of the proposed framework, and it also means that the sparse-scribble results attribute performance to the complete correction-aware formulation rather than isolating a single component.”

May combine two claims.

The proposed memory keeps the latest positive and negative instructions as the active instructions and records revised locations in a separate channel.

“The proposed memory keeps the latest positive and negative instructions as the active instructions and records revised locations in a separate channel.”

May combine two claims.

The model can use a small number of informative 2D corrections without requiring dense volume-wide scribbles.

“Together with Table 2, these results indicate that the model can use a small number of informative 2D corrections without requiring dense volume-wide scribbles.”

May combine two claims.

PRISM is evaluated under the same sparse test-time slice budgets, but its released checkpoint was trained with dense scribble assumptions, whereas the proposed model is trained with a mixture of dense and sparse informative-slice prompts.

“PRISM is evaluated under the same sparse test-time slice budgets, but its released checkpoint was trained with dense scribble assumptions, whereas the proposed model is trained with a mixture of dense and sparse informative-slice prompts.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The experiments use a single split and a single evaluation seed.

“Finally, the experiments use a single split and a single evaluation seed.”

May combine two claims.

The proposed sparse-prompt modeling and latest-write memory are not tied to a specific decoder in principle; they operate on prompt coordinates, the previous mask, and the refinement input.

“The proposed sparse-prompt modeling and latest-write memory are not tied to a specific decoder in principle; they operate on prompt coordinates, the previous mask, and the refinement input.”

Classified as probably not a substantial claim.

The implementation keeps the promptable backbone, prompt family, data splits, and simulation protocol as close as possible to PRISM so that the prompt process becomes the main variable under study.

“The implementation keeps the promptable backbone, prompt family, data splits, and simulation protocol as close as possible to PRISM so that the prompt process becomes the main variable under study.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

They could therefore be incorporated into other promptable 3D segmentation systems, although this paper evaluates them within a PRISM-compatible implementation to keep the comparison controlled.

“They could therefore be incorporated into other promptable 3D segmentation systems, although this paper evaluates them within a PRISM-compatible implementation to keep the comparison controlled.”