Pith. sign in

REVIEW 5 major objections 5 minor 29 references

Uncertainty-aware Diffusion and Reinforcement Learning for Joint Plane Localization and Anomaly Diagnosis in 3D Ultrasound

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Automated coronal-plane localization and seven-class uterine anomaly diagnosis can be performed jointly by a single diffusion-plus-RL framework, reaching 94.58% accuracy and 0.9989 AUC.

desk verdict A capable engineering paper whose new uncertainty mechanism needs a direct validation experiment; the L2 distance on spherical coordinates in Eq. (4) is a real technical concern, but the paper deserves a serious referee. read the letter →

arxiv 2506.23538 v2 pith:PIVNOLU4 submitted 2025-06-30 cs.CV cs.AIcs.LG

Yuhao Huang , Yueyue Xu , Haoran Dou , Jiaxiao Deng , Xin Yang , Hongyu Zheng , Dong Ni This is my paper
classification cs.CVcs.AIcs.LG
keywords congenitaluterineanomaly3Dultrasoundcoronalplanelocalizationdiffusionmodelreinforcementlearninguncertainty-awareclassificationkeyslicesummarytext-conditioneddenoising
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

Congenital uterine anomalies (CUAs) are typically diagnosed by finding the coronal plane in a 3D ultrasound volume and reading the morphology; both steps are currently manual and operator-dependent. The paper claims that a single framework can do both: a text-conditioned diffusion model localizes the coronal plane, an RL agent picks the most informative slices from the denoising trajectory for classification, and a text-driven uncertainty score reweights the classifier output. On a private dataset of 677 volumes spanning seven uterine categories, the full method reaches 94.58% accuracy, 95.21% F1, and 0.9989 AUC, with significant gains over 2D/2.5D, video, and 3D baselines. If correct, this turns two manual, biased steps into one automated pass that also returns the standard plane as an interpretable intermediate.

What carries the argument

The load-bearing mechanism is a text-conditioned denoising diffusion model whose plane prediction is parameterized in spherical tangent coordinates. Three condition sources—whole-volume features, current-plane features, and text embeddings from a pretrained biomedical text encoder—are merged through Sigmoid-gated adaptive weights that shift from global to local guidance as denoising proceeds. Along the trajectory, a REINFORCE-trained agent with a Bi-LSTM policy decides which slices to keep, balancing redundancy (cosine similarity), anatomical information (classification loss), and size (a penalty over five slices). The uncertainty module computes, for each candidate anomaly label, the spread of K denoised plane parameters produced under that label's text prompt, and uses the inverse of that spread to multiply the classifier probability before renormalizing.

What would settle it

Measure, on the held-out test set, the standalone accuracy of the coarse prediction $\arg\min_c S_{\mathrm{unc}}(c)$ and the correlation between $S_{\mathrm{unc}}$ and prediction error. If the arg-min class is no better than chance, or if the true category does not yield lower uncertainty than other categories, the uncertainty reweighting step is not responsible for the reported gains.

Watch

Extended reading notes

Core claim

The central claim is that plane localization and CUA diagnosis can be solved as one conditional denoising problem, and that diagnosis improves when the classifier is combined with the diffusion process itself. The plane is represented as a tangent point in spherical coordinates, denoised under adaptive weighting of volume, plane, and text conditions; each denoising trajectory yields a sequence of slices, and an RL agent selects a compact key-slice summary whose features feed the classifier. At test time, category-specific text prompts are run through the same denoiser, and the variance of the K predicted plane parameters under each prompt provides an uncertainty score: the prompt with lowest variance is the coarse diagnosis, and its normalized inverse uncertainty reweights the classifier probability. The paper reports this combined system dominates all compared localization and classification baselines, with statistically significant gains.

Load-bearing premise

The uncertainty adjustment rests on the premise that a text prompt naming the true anomaly makes the diffusion model's plane predictions more consistent across random restarts; the paper does not directly verify that this consistency tracks correctness.

Editorial extensions

If this is right

  • The same diffusion trajectory that localizes the plane also feeds the classifier, so the two clinical tasks can run in one pass rather than as separate models.
  • The RL slice summary is shown to improve classification across all tested localization methods, implying it is a portable add-on rather than a demand of this specific diffusion model.
  • Text-driven uncertainty requires no additional annotation, so the recalibration step can be appended to any diffusion-based locator that accepts category prompts.
  • Adaptive condition weighting separates the roles of global and local information over time, suggesting that coarse-to-fine guidance is sufficient for standard-plane search.

Reading between the lines

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

  • A natural standalone test would treat the arg-min uncertainty prediction as a zero-shot classifier: if it disagrees worse than chance, the uncertainty signal is not measuring what the paper assumes.
  • If the uncertainty score truly tracks correctness, it could be used for selective prediction, flagging low-confidence volumes for human review rather than forcing a single answer.
  • The key-slice selection idea transfers beyond this task to any iterative localization method that emits an ordered set of candidate planes, such as fetal or cardiac ultrasound.
  • Because the framework returns a standard plane alongside the diagnosis, it could support semi-automatic documentation in clinical workflows rather than only an end label.
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.

Referee Report

5 major / 5 minor

Summary. This manuscript presents a joint framework for coronal standard-plane localization and seven-class congenital uterine anomaly (CUA) classification from 3D ultrasound volumes. The method combines a diffusion-based plane regressor with adaptive weighting of plane, volume, and text conditions; a REINFORCE-trained module that selects key slices from the denoising trajectory; and a text-conditioned uncertainty score that reweights the classification probabilities at inference. Experiments on a private dataset of 677 volumes compare against five localization baselines and thirteen classification baselines, with ablations and statistical tests. The authors report strong gains: e.g., F1 95.21% and AUC 0.9989 for the full model, and significantly improved localization over prior diffusion work.

Significance. If the results hold, the system would be an interesting step toward integrated plane localization and diagnosis in 3D US, matching the clinical workflow of coronal-plane reconstruction followed by morphological assessment. The paper's strengths include a relatively large private dataset with seven categories, a broad set of localization and classification baselines, pairwise statistical tests, and public code. The main risk is the uncertainty module: its core assumption is not directly validated, and the proposed score is defined on a coordinate representation for which Euclidean distance is not a proper metric. The ablation table also contains an unexplained performance drop. These issues are fixable but currently prevent full confidence in the central claim.

major comments (5)
  1. [Section 2.3, Eq. (4)-(5)] The entire uncertainty-aware adjustment rests on the assumption that the correct text condition yields lower dispersion among the K=8 denoised plane parameters, yet the paper provides no direct evidence for this correlation. The only support is the end-to-end accuracy gain (Table 2, last two rows), which conflates the uncertainty signal with the other components. Please report the distribution of Sunc for correct versus incorrect conditions, the rank correlation between Sunc and localization/classification correctness, or an ablation in which uncertainty scores are replaced by random permutations. Moreover, because the training prompts include the ground-truth class label (Section 2.1), the observed uncertainty differences may be confounded by per-class training frequency or category-specific parameter distributions rather than by semantic correctness; this confound should be addressed. Without such evidence, argmin_c Sunc and the reweighting in Eq. (5) are not established.
  2. [Eq. (4)] Sunc is computed as a sum of squared Euclidean distances between p_k and the mean in the raw parameter space p=(r, eta, theta), where eta and theta are angular coordinates. Euclidean distance on angles is not invariant to coordinate origin/rotation, has a branch-cut discontinuity, and is sensitive to the arbitrary scaling between r and the angular components, so the score may reflect parameterization artifacts rather than localization or classification confidence. A metric that respects the spherical geometry (e.g., chordal or geodesic distance on the tangent space) should be used, or the parameters should be converted to Cartesian coordinates before computing dispersion. The summation also uses index i on the left of Eq. (4) but k inside; this should be corrected.
  3. [Table 2] The ablation is missing a controlled SS-only condition, and the AP+GF row shows a dramatic, unexplained F1 drop from 91.96 (AP only) to 71.46 when global volume features are added to the all-plane input. Since the final model combines SS, GF, and UA, this non-monotonicity makes it hard to attribute the final gains to the proposed components and suggests the global-feature fusion may be unstable. Please add the missing ablation cells, report multiple seeds, and provide a discussion of this drop.
  4. [Section 2.2, Eq. (2)] The reward is described as unsupervised, but R_cls is the cross-entropy between the summary-based predictions and the ground-truth class labels y_c; the RL policy is therefore trained with a label-dependent reward. This is a learned, label-based reward, not an unsupervised one. The terminology should be corrected, and the possibility that the slice-summary gains partly reflect overfitting to the training labels should be addressed, for example by evaluating the selected slices on a held-out set or with class-balanced rewards.
  5. [Section 3, Table 2] Classification metrics are reported as point estimates on a single split, without confidence intervals or per-class results for the highly imbalanced seven-class problem (e.g., bicornuate n=5, T-shaped n=15, didelphys n=19). The chi-square p-value only supports overall accuracy differences; it does not establish improvements in precision, recall, or F1. Please report seed-based variability and per-class metrics, and specify whether Pre/Rec/F1 are macro-averaged.
minor comments (5)
  1. [References] The Pair annotation software is cited as reference [18], but [18] describes a GAN for ultrasound image synthesis, not an annotation tool; please cite the correct software or describe it explicitly.
  2. [Section 2.3] The text prompts are written as 'This is a {U} in 3D ultrasound', but the actual class names and the exact prompt templates used for training and inference are not listed; please provide them for reproducibility.
  3. [Table 2] The columns AP, PP, and SS are mutually exclusive input choices, so presenting them as independent checkboxes is confusing; a single 'classification input' column or explicit row labels would be clearer.
  4. [Section 2.2, Eq. (3)] The L2 regularization term in the RL update (eta * sum of zeta^2) is not introduced; please define eta and state whether it is a weight-decay coefficient.
  5. [Figures 2 and 3] The figures are not fully self-contained; please define the color/selection conventions in the captions (e.g., green vs gray in Fig. 2) and identify the meaning of each column in Fig. 3.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the joint localization-diagnosis derivation is empirically self-contained; the uncertainty adjustment is an explicit heuristic and the self-citations are ordinary methodological inheritance.

full rationale

The paper's derivation chain does not reduce to its own inputs. The diffusion plane localizer is built on the authors' prior work [7] for the tangent-point spherical parameterization and diffusion formulation, but this is a normal, externally falsifiable methodological inheritance rather than a load-bearing citation of an unverified theorem, and the new contributions (adaptive condition weighting, RL slice summary, uncertainty adjustment) are defined by explicit equations and evaluated on held-out splits against external baselines. The uncertainty-aware strategy in Sec. 2.3 is presented as an assumption, not a definition: 'we assume that giving the correct text condition will bring less uncertainty,' and Eq. (4)-(5) are fixed heuristic formulas with no parameter fitted to the test labels, so the resulting 'coarse prediction' and probability adjustment are not statistically forced. Even if the assumption is unvalidated or Eq. (4)'s L2 distance on spherical coordinates is a poor metric, those are correctness risks, not cases where a claimed prediction is equivalent to its input by construction. Tables 1-2 compare against external and baseline methods, and the reported improvements are not built into the definitions of the metrics. No circular step is present.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

No new physical entities are proposed. The method relies on learned condition weights, hand-set RL reward hyperparameters, and an explicit but unvalidated assumption about text-conditioned uncertainty. These are the main elements the reader must accept on faith.

free parameters (5)
  • Adaptive condition weights omega_v, omega_p, omega_t = learned during training
    Linear projections with Sigmoid map concatenated features to weights for volume, plane, and text conditions in the diffusion denoiser (Section 2.1). These are fitted to training data.
  • RL reward penalty weight alpha = not reported
    Scales the size penalty term in Eq. (2); chosen by hand with no sensitivity analysis.
  • RL reward growth rate gamma = not reported
    Controls how sharply the size penalty grows when |S| exceeds S_max in Eq. (2); chosen by hand.
  • S_max (maximum slice summary size) = 5
    Hard limit on the number of selected slices in the RL summary; set in Section 2.2 with no justification.
  • K (number of stochastic denoising runs for uncertainty) = 8
    Number of random initial plane parameters used to compute the uncertainty score in Section 2.3 and Eq. (4); chosen by hand.
assumptions (4)
  • domain assumption The standard coronal plane annotated by sonographers is the correct target for CUA diagnosis.
    Underpins the plane localization task and the ground-truth labels; stated in Section 1 and the experimental setup (Section 3).
  • ad hoc to paper Correct text conditions produce lower variance in the predicted plane parameters than incorrect ones.
    Stated explicitly in Section 2.3: "we assume that giving the correct text condition will bring less uncertainty." This assumption is load-bearing for the uncertainty adjustment and is not separately validated.
  • ad hoc to paper The RL reward in Eq. (2) balances redundancy, classification accuracy, and summary size, and is a good proxy for selecting diagnostically relevant slices.
    No supervised key-slice labels exist; the reward design is heuristic (Section 2.2).
  • standard math The diffusion plane parameterization and noise schedule from [7,20] are appropriate for this task.
    Adopted from prior work (Section 2.1); not re-derived here.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Uncertainty-aware Diffusion and Reinforcement Learning for Joint Plane Localization and Anomaly Diagnosis in 3D Ultrasound." pith.science (2026). https://pith.science/paper/PIVNOLU4

@misc{pith2026250623538,
  author       = {Pith},
  title        = {Pith review of: Uncertainty-aware Diffusion and Reinforcement Learning for Joint Plane Localization and Anomaly Diagnosis in 3D Ultrasound},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PIVNOLU4}},
  note         = {Machine review of arXiv:2506.23538}
}
read the original abstract

Congenital uterine anomalies (CUAs) can lead to infertility, miscarriage, preterm birth, and an increased risk of pregnancy complications. Compared to traditional 2D ultrasound (US), 3D US can reconstruct the coronal plane, providing a clear visualization of the uterine morphology for assessing CUAs accurately. In this paper, we propose an intelligent system for simultaneous automated plane localization and CUA diagnosis. Our highlights are: 1) we develop a denoising diffusion model with local (plane) and global (volume/text) guidance, using an adaptive weighting strategy to optimize attention allocation to different conditions; 2) we introduce a reinforcement learning-based framework with unsupervised rewards to extract the key slice summary from redundant sequences, fully integrating information across multiple planes to reduce learning difficulty; 3) we provide text-driven uncertainty modeling for coarse prediction, and leverage it to adjust the classification probability for overall performance improvement. Extensive experiments on a large 3D uterine US dataset show the efficacy of our method, in terms of plane localization and CUA diagnosis. Code is available at https://github.com/yuhoo0302/CUA-US.

Figures

Figures reproduced from arXiv: 2506.23538 by the authors.

Figure 1
Figure 1. Overview of our proposed framework. plane refinement conditioned on multi-scale guidance (i.e., plane, volume and text). In stage 2, we leverage the RL strategy to select the key slices from the above iterative process to build the slice summary and boost model learning. During inference, we obtain the text-driven uncertainty scores to adjust the original classification probabilities, finally improving the overall p… view at source ↗
Figure 2
Figure 2. Typical summary examples with uterine types in the top-left corners. Below are two summaries with index selection probabilities: green (select) and gray (skip). of two planes, and Dis (mm) denotes the difference between their Euclidean distances towards the volume origin. Details for SSIM and NCC refer to [7]. For classification, Accuracy (Acc, %), Precision (Pre, %), Recall (Rec, %), F1-score (F1, %), and AUC were … view at source ↗
Figure 3
Figure 3. Visualization results. c1: predicted planes with GTs (yellow). c2: denoising curves of condition weights at different time steps. c3-5: original, uncertainty and ad￾justed probabilities with predicted classification results (c: column). slice summary as classification inputs, respectively. GF means the global vol￾umetric features, and UA denotes the uncertainty-aware strategy. We observe that even with GF, using onl… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 19 canonical work pages

  1. [1]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Alansary, A., et al.: Automatic view planning with multi-scale deep reinforcement learning agents. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 277–285. Springer (2018)

    work page 2018

  2. [2]

    In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299–6308 (2017)

    2017

  3. [3]

    Human reproduction update17(6), 761–771 (2011)

    Chan,Y.Y., Jayaprakasan, K., Zamora, J.,Thornton, J.G., etal.: The prevalence of congenital uterine anomalies in unselected and high-risk populations: a systematic review. Human reproduction update17(6), 761–771 (2011)

    work page 2011

  4. [4]

    arXiv preprint arXiv:1904.00625 (2019)

    Chen, S., Ma, K., Zheng, Y.: Med3d: Transfer learning for 3d medical image anal- ysis. arXiv preprint arXiv:1904.00625 (2019)

    arXiv 2019

  5. [5]

    Ad- vances in Neural Information Processing Systems36(2024)

    Clark, K., Jaini, P.: Text-to-image diffusion models are zero shot classifiers. Ad- vances in Neural Information Processing Systems36(2024)

    work page 2024

  6. [6]

    arXiv preprint arXiv:2010.11929 (2020)

    Dosovitskiy, A., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

    arXiv 2020

  7. [7]

    Computer Methods and Programs in Biomedicine p

    Dou, H., Huang, Y., Huang, Y., Yang, X., Zhen, C., Zhang, Y., Xiong, Y., Huang, W., Ni, D.: Standard plane localization using denoising diffusion model with multi- scale guidance. Computer Methods and Programs in Biomedicine p. 108619 (2025)

    2025

  8. [8]

    arXiv preprint arXiv:2502.03687 (2025)

    Favero, G.M., Saremi, P., Kaczmarek, E., Nichyporuk, B., Arbel, T.: Conditional diffusion models are medical image classifiers that provide explainability and un- certainty for free. arXiv preprint arXiv:2502.03687 (2025)

    arXiv 2025

Show all 29 references
  1. [9]

    In: ISBI

    Gong, H., et al.: nnmamba: 3d biomedical image segmentation, classification and landmark detection with state space model. In: ISBI. pp. 1–5. IEEE (2025)

    2025

  2. [10]

    In: CVPR

    He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770–778 (2016) 10 Y. Huang et al

    2016

  3. [11]

    In: CVPR

    Huang, G., Liu, Z., et al.: Densely connected convolutional networks. In: CVPR. pp. 4700–4708 (2017)

    2017

  4. [12]

    Medical Image Analysis80, 102490 (2022)

    Huang, R., Ying, Q., et al.: Extracting keyframes of breast ultrasound video using deep reinforcement learning. Medical Image Analysis80, 102490 (2022)

    2022

  5. [13]

    In: 3D Ultrasound, pp

    Huang, Y., Zou, Y., Dou, H., Huang, X., Yang, X., Ni, D.: Localizing standard plane in 3d fetal ultrasound. In: 3D Ultrasound, pp. 239–269. CRC Press (2023)

    2023

  6. [14]

    In: CVPR

    Jang, J., et al.: M3t: three-dimensional medical image classifier using multi-plane and multi-slice transformer. In: CVPR. pp. 20718–20729 (2022)

    2022

  7. [15]

    Archives of Gynecology and Obstetrics299, 779–789 (2019)

    Kougioumtsidou, A., Mikos, T., Grimbizis, G.F., et al.: Three-dimensional ultra- sound in the diagnosis and the classification of congenital uterine anomalies using the eshre/esge classification: a diagnostic accuracy study. Archives of Gynecology and Obstetrics299, 779–789 (2019)

    2019

  8. [16]

    arXiv preprint arXiv:2211.09552 (2022)

    Li, K., Wang, Y., He, Y., Li, Y., et al.: Uniformerv2: Spatiotemporal learning by arming image vits with video uniformer. arXiv preprint arXiv:2211.09552 (2022)

    2022 arXiv

  9. [17]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Li, Y., Khanal, B., Hou, B., Alansary, A., Cerrolaza, J.J., et al.: Standard plane detection in 3d fetal ultrasound using an iterative transformation network. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 392–400. Springer (2018)

    2018

  10. [18]

    Medical image analysis79, 102461 (2022)

    Liang, J., Yang, X., Huang, Y., Li, H., He, S., Hu, X., Chen, Z., Xue, W., Cheng, J., Ni, D.: Sketch guided and progressive growing gan for realistic and editable ultrasound image synthesis. Medical image analysis79, 102461 (2022)

    2022

  11. [19]

    In: CVPR

    Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swin trans- former. In: CVPR. pp. 3202–3211 (2022)

    2022

  12. [20]

    arXiv preprint arXiv:2010.02502 (2020)

    Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)

    2020 arXiv

  13. [21]

    In: International Conference on Medical Im- age Computing and Computer-Assisted Intervention

    Sun, A., Zhang, Z., et al.: Boosting breast ultrasound video classification by the guidance of keyframe feature centers. In: International Conference on Medical Im- age Computing and Computer-Assisted Intervention. pp. 441–451. Springer (2023)

    2023

  14. [22]

    Wang, Y., Li, Z., et al.: Key-frame guided network for thyroid nodule recognition usingultrasoundvideos.In:InternationalConferenceonMedicalImageComputing and Computer-Assisted Intervention. pp. 238–247. Springer (2022)

    2022

  15. [23]

    Machine learning8, 229–256 (1992)

    Williams, R.J.: Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning8, 229–256 (1992)

    1992

  16. [24]

    Scientific Data10(1), 41 (2023)

    Yang, J., Shi, R., et al.: Medmnist v2-a large-scale lightweight benchmark for 2d and 3d biomedical image classification. Scientific Data10(1), 41 (2023)

    2023

  17. [25]

    IEEE Transactions on Medical Imaging40(7), 1950–1961 (2021)

    Yang, X., Dou, H., Huang, R., Xue, W., Huang, Y., et al.: Agent with warm start and adaptive dynamic termination for plane localization in 3d ultrasound. IEEE Transactions on Medical Imaging40(7), 1950–1961 (2021)

    2021

  18. [26]

    Medical Image Analysis72, 102119 (2021)

    Yang, X., Huang, Y., Huang, R., Dou, H., Li, R., Qian, J., Huang, X., Shi, W., Chen, C., Zhang, Y., et al.: Searching collaborative agents for multi-plane local- ization in 3d ultrasound. Medical Image Analysis72, 102119 (2021)

    2021

  19. [27]

    In: 17th International Workshop on Breast Imaging (IWBI 2024)

    Yang, Z., Fan, T., Smedby, Ö., Moreno, R.: 3d breast ultrasound image classifica- tion using 2.5 d deep learning. In: 17th International Workshop on Breast Imaging (IWBI 2024). vol. 13174, pp. 443–449. SPIE (2024)

    2024

  20. [28]

    arXiv preprint arXiv:2303.00915 (2023)

    Zhang, S., Xu, Y., Usuyama, N., Xu, H., Bagga, J., Tinn, R., Preston, S., et al.: Biomedclip: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs. arXiv preprint arXiv:2303.00915 (2023)

    2023 arXiv

  21. [29]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Zou, Y., Dou, H., Huang, Y., Yang, X., Qian, J., Zhen, C., et al.: Agent with tangent-based formulation and anatomical perception for standard plane localiza- tion in 3d ultrasound. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp....

    2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.