Automatic LV Localization and Short-Axis Plane Estimation from Arbitrary CMR Slice

Parker Martin; Yixuan Liu; Yi Yu; Yuan Xue; Yuchi Han; Zhenyu Bu; Ziyu Zhang

REVIEW 4 major objections 5 minor 29 references

Automatic LV Localization and Short-Axis Plane Estimation from Arbitrary CMR Slice

T0 review · 4 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 The paper claims that a single arbitrary 2D CMR slice is enough to localize the left ventricle and recover the full 3D orientation of the short-axis plane, using a continuous unit-energy embedding.

desk verdict New single-slice CMR orientation task with a sound method, but the 'arbitrary slice' claim outruns the benchmark's LV-centered slice distribution. read the letter →

arxiv 2608.00145 v1 pith:YUAAXRRY submitted 2026-07-31 eess.IV cs.CV

Yi Yu , Yixuan Liu , Ziyu Zhang , Parker Martin , Zhenyu Bu , Yuchi Han , Yuan Xue This is my paper

classification eess.IVcs.CV

keywords leftventriclelocalizationCMRorientationestimationshort-axisplanepolar-coupledcircularembeddingorientedobjectdetectionsingle-sliceregressioncardiacMRIrotatedboundingbox

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

The paper is trying to establish a new task: from one arbitrary 2D cardiac MRI slice, a network can both locate the left ventricle and estimate the 3D direction of the short-axis plane. It argues that existing orientation representations — view classification, direct angle regression, IoU/Gaussian overlap, and the phase-shifting coder — fail on this task due to 180° ambiguity, boundary discontinuity, or instability near the poles. The proposed Polar-Coupled Circular (PCC) embedding encodes the plane normal as a unit-norm circular-harmonic vector that is continuous, unambiguous, and geometrically consistent. If right, this would allow SAX plane prescription and spatially aware analysis from a single slice rather than multi-view intersection. Reported results are an average mIoU of 86.18% and an average angle deviation of 3.39° across four datasets.

What carries the argument

The Polar-Coupled Circular (PCC) embedding: for a unit normal n, the embedding coefficients are p_n = sqrt(2/(N-1)) * sqrt(1-n_z^2) * cos(theta + 2πn/(N-1)) plus p_{N-1}=n_z. The key property is that all coefficients together have unit Euclidean norm, and the azimuthal components are scaled by sqrt(1-n_z^2), so as the plane normal approaches the poles the azimuthal information smoothly fades out. This gives a continuous, one-to-one, boundary-free parameterization of the semi-sphere, which lets a standard FCOS-style detector regress orientation with ordinary L1 loss.

What would settle it

Benchmark the trained model on a new cohort of, say, 50 patients whose SAX normals are taken from DICOM metadata of clinically acquired single-slice localizers. If the average angle deviation on those real slices rises substantially above the reported 3.39°, the single-slice recoverability claim would be refuted. A second probe: feed the model slices that do not intersect the LV mask; if it confidently predicts an orientation rather than reporting absence, the localization-orientation coupling is incomplete.

Watch

Extended reading notes

Core claim

The paper's central claim is that 3D SAX orientation can be regressed directly from a single 2D CMR slice, and that the PCC embedding makes this regression stable in ways that previous rotation representations are not. The ground-truth SAX unit normal is encoded by the PCC map into N circular-harmonic coefficients with unit-energy norm; the network regresses these coefficients, and a linear projection decodes them back to a Cartesian normal. The authors show this beats four adapted orientation-aware detectors in both orientation accuracy and localization consistency, with lower standard deviation, across four datasets spanning different centers, vendors, and cardiac phases.

Load-bearing premise

The assumption that a single 2D CMR slice containing part of the left ventricle carries enough information to determine the full 3D orientation of the short-axis plane, with the random planes sampled through LV masks in segmentation volumes treated as representative of arbitrary clinical slices.

Editorial extensions

If this is right

SAX plane prescription could be performed from a single localizer slice, removing the need for multi-slice geometric intersection.
Orientation predictions over the full semi-sphere are unambiguous: 180° flips that plague IoU/Gaussian-based methods are avoided.
The synthetic slicing pipeline turns commonly available 3D segmentation datasets into large-scale orientation training data without manual annotation.
The method generalizes across multi-center and multi-vendor datasets and across cardiac phases (In-House 30-frame annotation), suggesting phase-invariant orientation cues.

Reading between the lines

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

The same unit-energy circular-harmonic encoding applies to any task where a 3D plane normal must be regressed from a 2D view — e.g., other standard cardiac planes (2CH/4CH) or fetal/neonatal imaging — so the mechanism is not LV-specific.
If truly arbitrary clinical slices include planes that miss the ventricle, the detector would need an explicit 'no LV' output; the current benchmark guarantees LV visibility by construction, so rejection behavior remains untested.
The small real-world validation (four volunteers) leaves open how the performance transfers to patients with pathology or unusual anatomy; a larger prospective study would close that gap.

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

The task here is genuinely new: regress the continuous 3D normal of the LV short-axis plane from a single 2D CMR slice, while localizing the LV with a rotated box. Prior work stops at discrete view classification or needs multiple slices. The PCC embedding is a clean extension of circular-harmonic encodings to 3D with a unit-energy constraint, and their adaptation of four detection baselines (FCOS3D, RIUo3D, GWD3D, PSC) gives a useful picture of where those fail. The benchmark pipeline that resamples planes from existing segmentation volumes is practical and will be reused.

That said, the central claim is oversold. Every training and test slice is generated by sampling a plane through a randomly chosen point inside the LV mask, so the ventricle is always visible and roughly centered. The evaluation therefore concerns the conditional distribution 'slice contains LV', not 'arbitrary CMR slice'. The four-volunteer real-world prescription test is encouraging but too small to close the gap. The paper also selects N=4 and w_emb=2 by ablation on V2S-Real, which then appears in the main test table — textbook test-set tuning, though the ablation differences are small enough that the conclusion likely survives.

Two smaller points: Table 1 shows the same FCOS3D numbers for M&Ms-2 and In-House, which is almost certainly a copy-paste error. And Eq. (1), if read with the square root over the full factor, does yield unit norm (the sum of (N-1) equally spaced cos^2 terms is (N-1)/2), so I would not flag that as an error.

Overall, the core idea is sound, the method is sensible, and the benchmark is a real contribution. The paper deserves a serious referee. My main request would be to either soften the 'arbitrary' language or add an evaluation on slices that miss the LV, and to fix the minor editorial issues.

Referee Report

4 major / 5 minor

Summary. The paper introduces a new task: given a single 2D CMR slice, jointly localize the left ventricle (as a rotated 2D box) and estimate the 3D orientation of the short-axis plane. It adapts four orientation-aware detection methods (RIoU3D, FCOS3D, GWD3D, PSC) to this setting, analyzes their limitations, and proposes Polar-Coupled Circular (PCC) embedding: a circular-harmonic representation of the SAX normal coupled with the polar component. A benchmark is constructed by randomly slicing 3D CMR segmentation volumes through the LV mask, and experiments on ACDC, M&Ms-2, In-House, and V2S-Real report average mIoU 86.18%, average angle deviation 3.39°, and AP75 87.96%. A real-world SAX plane prescription experiment on four volunteers is also reported.

Significance. If the results hold, the paper opens a useful direction for automated cardiac plane prescription and cross-view analysis: estimating a continuous 3D orientation from a single slice is more general than discrete view classification and does not require multi-slice geometric intersection. The PCC embedding is geometrically motivated, and the paper includes reproducible assets: code is released, the benchmark generation pipeline is described in detail, and orientation labels are derived from external known SAX geometry, so there is no circularity in the central evaluation. The comparison with four adapted baselines is also informative. The main risks are the mismatch between the benchmark slice-sampling distribution and truly arbitrary clinical slices, the very small real-world validation, and a typographical issue in Eq. (1) that, as printed, breaks the central unit-energy property.

major comments (4)

[§3.1–3.2, Eq. (1)–(3)] As typeset, Eq. (1) defines p_n = sqrt(2/(N-1)) (1 - n_z^2) cos(θ + 2πn/(N-1)). With this definition, ||u||_2^2 = (1-n_z^2)^2 + n_z^2, which is not identically 1, contradicting the claimed unit-energy constraint. The reconstruction in Eq. (3) then recovers (1-n_z^2)(cos θ, sin θ) rather than (n_x, n_y). The missing factor appears to be sqrt(1-n_z^2); with that factor, the sum of squares of the circular components equals 1-n_z^2 and the full embedding has unit norm, and Eq. (3) correctly recovers the normal. Please correct Eq. (1) and verify the reconstruction, or explicitly justify a different intended definition.
[§3.4 and Fig. 5] The benchmark is built by first sampling a point P inside the LV mask and then sampling a plane through P. This guarantees that every training and test slice intersects the LV, so the model is evaluated only on the conditional distribution P(slice | LV is visible). The title, abstract, and introduction claim 'arbitrary CMR slice,' but non-LV slices, severely partial LV slices, and clinically typical localizer views are excluded. The real-world validation in Fig. 5 uses only four volunteers and does not include negative or non-LV slices. Therefore Table 1's numbers cannot be read as evidence for the general 'arbitrary slice' claim. I ask for either (i) additional evaluation on a retrospectively collected set of true clinical slices, including slices that do not intersect the LV, or (ii) a clearly qualified statement that the task is single-slice SAX-plane estimation conditional on LV pres
[Table 1, In-House row] The In-House row for FCOS3D is identical to the M&Ms-2 row (4.94±3.84, 84.22, 99.78, 81.44). This appears to be a copy/paste error. Since Table 1 is the central quantitative evidence for the cross-dataset improvement, this duplicated row must be corrected or recomputed, and any averaged performance numbers that depend on it must be updated.
[§4, Fig. 5c] The multi-view comparison is based on only four volunteers, and the figure reports only aggregate mean±std. No per-volunteer measurements, number of localizer slices, acquisition details, or statistical tests are provided. With n=4, the claim of being 'on par with' or better than multi-view methods is not established. Please report the underlying individual measurements and a statistical comparison, or present this as a preliminary pilot result.

minor comments (5)

[§3.2] The decoding step does not state whether the recovered normal is renormalized after projection. Once Eq. (1) is fixed this may be unnecessary, but the paper should state explicitly whether any post-processing normalization is applied.
[§2 / Table 1] For RIoU3D and GWD3D, the 180° rotational ambiguity means the network can predict either n or -n. Reporting angle deviation with this sign ambiguity conflates 'reversed prediction' with geometric error; consider also reporting the minimal axis-angle deviation for these baselines to clarify the failure mode.
[§3.4] The text says '20 slices are sampled from each mask-annotated 3D CMR volume,' but then V2S-Real is described as containing 2,000 frames. It would be helpful to state the resulting total number of training/test slices per dataset and whether slices from the same volume appear in both training and test sets.
[§4] The In-House subset description says 188 scans with 95 for testing, but does not explain how the training split is formed or whether there is patient overlap between training and test. Please clarify this for reproducibility.
[Throughout] The term 'arbitrary CMR slice' is used much more broadly than the actual experimental setting. I recommend consistently qualifying the claim as 'arbitrary slice intersecting the LV' unless the broader evaluation is added.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: PCC is a supervised invertible encoding; benchmark labels are external geometry; self-citations are baselines, not evidence.

full rationale

The central derivation is a supervised regression from a 2D slice to a 3D orientation. Training labels are generated by transforming the known SAX direction into the slice coordinate system (Sec. 3.4), and the network is supervised with L1 loss against the PCC-encoded normal (Eqs. 4-5). The encoding Φ_PCC (Eq. 1) and its inverse (Eq. 3) form a bijective representation; the paper does not use the inverse to produce labels from network outputs, so the prediction is not equivalent to the input by construction. The benchmark's stochastic slicing guarantees LV visibility (point P lies inside the LV mask), which limits external validity for truly arbitrary clinical slices, but does not make the evaluation circular—the orientation and bounding-box labels are independent geometric facts derived from the known 3D volume and SAX direction. Self-authored references [24,25] (PSC) are used as comparison baselines, and the claimed PCC improvement is measured on held-out test sets, not imported from those citations. The real-world validation with 4 volunteers (Fig. 5) is small but is an external comparison, not a circular step. No fitted parameter is renamed as a prediction; no uniqueness theorem is borrowed from the authors' prior work. Therefore the circularity score is 0.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper's central empirical claim rests on a synthetic benchmark built from segmentation masks. It introduces no new physical entities; its free parameters are two hyperparameters tuned on one dataset that also appears in the main evaluation.

free parameters (2)

PCC embedding length N = 4
Selected by ablation on V2S-Real (Table 2); N=4 gives best angle deviation/AP75. This is a hyperparameter tuned on the same dataset reported in main results.
Embedding loss weight w_emb = 2
Selected by ablation on V2S-Real (Table 2); w_emb=2 gives best trade-off. Same test-set tuning caveat.

assumptions (4)

domain assumption A single 2D CMR slice's appearance provides enough information to uniquely recover the 3D SAX normal vector, up to inherent pole degeneracy.
The entire task rests on this learnability/identifiability assumption; the paper provides empirical evidence but no theoretical identifiability analysis (Sec. 1, Sec. 4).
domain assumption The known SAX direction from volumetric segmentation datasets is a reliable ground-truth 3D orientation, and transforming it into random slicing coordinates yields correct labels.
Sec. 3.4 uses 'SAX direction (which is known for SAX stacks and manually annotated otherwise)' as ground truth; errors in this assumption propagate directly into benchmark labels.
domain assumption Random planes intersecting the LV mask produce 2D slices representative of clinical CMR slices.
Sec. 3.4 samples planes through P inside the LV mask, guaranteeing LV visibility; this excludes slices without the LV and may not match real acquisition geometry.
standard math Unit-norm circular-harmonic encoding/decoding (Eqs. 1-3) is a correct continuous representation of the 3D unit sphere; standard Fourier projection math.
Orthogonal projection onto first-order cosine/sine bases recovers azimuth, assuming Eq. (1) is corrected to include sqrt(1-nz^2); as printed the unit-norm claim fails.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Automatic LV Localization and Short-Axis Plane Estimation from Arbitrary CMR Slice." pith.science (2026). https://pith.science/paper/YUAAXRRY

@misc{pith2026260800145,
  author       = {Pith},
  title        = {Pith review of: Automatic LV Localization and Short-Axis Plane Estimation from Arbitrary CMR Slice},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YUAAXRRY}},
  note         = {Machine review of arXiv:2608.00145}
}

read the original abstract

Accurate estimation of left ventricular (LV) orientation is essential for cardiac magnetic resonance (CMR) imaging and downstream analysis. Existing methods typically formulate orientation recognition as discrete view classification or rely on multi-slice geometric intersection, limiting their ability to model continuous 3D orientation and generalize across arbitrary slices. This work introduces a novel paradigm: Joint LV localization and 3D orientation estimation from a single CMR slice. To investigate this setting, representative orientation-aware detection frameworks are adapted to the CMR domain, and their limitations are analyzed. Upon that, we propose the Polar-Coupled Circular (PCC) embedding that provides a continuous and unambiguous orientation representation to address the limitations. Meanwhile, a scalable benchmark is constructed through automatic slice sampling from volumetric CMR segmentation datasets. Extensive experiments on four datasets demonstrate strong performance, achieving an average mIoU of 86.18% and an average angle deviation of 3.39{\deg}. This study establishes a new task setting for single-slice LV orientation modeling and provides a geometry-consistent framework for spatially informed CMR analysis. Code is available at https://github.com/yuyi1005/cmr-3d-ood.

Figures

Figures reproduced from arXiv: 2608.00145 by the authors.

**Figure 1.** Orientation recognition from a single image has achieved remarkable success in autonomous driving. However, in the CMR domain, orientation estimation still largely relies on traditional view classification or plane-intersection strategies. Existing research on SAX orientation recognition, however, largely follows two paradigms: 1) Some studies formulate it as a view classification problem, assigning images to predef… view at source ↗

**Figure 2.** Visualization of our work. Our model localizes the left ventricle with a rotated bounding box and estimates its 3D orientation (red arrows) using a single, arbitrary 2D slice. The second row displays some compared methods repurposed to our task. from images [16]. Representative methods include: 1) L1-based. Early methods directly regress 3D box parameters, including rotation angles, using L1 or smooth L1 losses. Fo… view at source ↗

**Figure 3.** The framework of our proposed model. based methods such as PSC alleviate 2D boundary problem but are not inherently designed for 3D orientation. Extending them with an additional z-channel leads to varying embedding magnitudes across different directions and instability near polar regions (e.g., z → 1), resulting in unreliable predictions. Overall, existing formulations work well in LV localization, but exhibit rep… view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Illustration of the data generation pipeline. Readily available 3D segmentation datasets are repurposed for our task setting, LV localization and oriented detection. Finally, with the weighted losses wclsLcls, wboxLbox, and wcenLcen defined by backbone FCOS detector, t…

**Figure 5.** Figure 5: Comparison on the real-world SAX plane prescription task. We compare our predicted SAX normal with manually prescribed planes, demonstrating performance on par with existing automatic multi-view plane prescription methods [15, 22, 8]. using [27]. The isotropic resoluti…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 1 linked inside Pith

[1]

European Heart Journal - Cardiovascular Imaging20(Supplement2), jez116.025 (2019)

Aung, N., Lee, A.M., Sanghvi, M.M., Fung, K., Paiva, J.M., Thomson, R.J., et al.: Automatic classification of cmr image sequences with convolutional neu- ral networks. European Heart Journal - Cardiovascular Imaging20(Supplement2), jez116.025 (2019)

2019
[2]

In: Medical Imaging 2011: Image Process- ing

Bağci, U., Udupa, J.K., Chen, X.: Orientation estimation of anatomical structures in medical images for object recognition. In: Medical Imaging 2011: Image Process- ing. vol. 7962, p. 79622L. International Society for Optics and Photonics (2011)

2011
[3]

Bernard, O., Lalande, A., Zotti, C., Cervenansky, F., Yang, X., Heng, P.A., et al.: Deep learning techniques for automatic mri cardiac multi-structures segmenta- tion and diagnosis: Is the problem solved? IEEE Transactions on Medical Imaging 37(11), 2514–2525 (2018)

2018
[4]

Radiology: Artificial Intelligence1(6), e180069 (2019)

Blansit, K., Retson, T., Masutani, E., Bahrami, N., Hsiao, A.: Deep learning–based prescription of cardiac mri planes. Radiology: Artificial Intelligence1(6), e180069 (2019)

2019
[5]

Journal of Magnetic Resonance Imaging63(3), 891–903 (2026)

Böttcher, B., Meinel, F.G., Deyerberg, K.K., Watzke, L.M., Manzke, M., Gorodezky,M.,etal.:Fullyautomatedplaneprescriptionincardiacmri:Aprospec- tive cohort study. Journal of Magnetic Resonance Imaging63(3), 891–903 (2026)

2026
[6]

IEEE Transactions on Medical Imaging40(12), 3543–3554 (2021)

Campello, V.M., Gkontra, P., Izquierdo, C., Martín-Isla, C., Sojoudi, A., Full, P.M., et al.: Multi-centre, multi-vendor and multi-disease cardiac segmentation: The m&ms challenge. IEEE Transactions on Medical Imaging40(12), 3543–3554 (2021)

2021
[7]

Clinical Imaging82, 121–126 (2022) 10 Y

Chauhan, D., Anyanwu, E., Goes, J., Besser, S.A., Anand, S., Madduri, R., et al.: Comparison of machine learning and deep learning for view identification from cardiac magnetic resonance images. Clinical Imaging82, 121–126 (2022) 10 Y. Yu et al

2022
[8]

Medical Physics49(1), 129–143 (2022)

Edalati, M., Zheng, Y., Watkins, M.P., Chen, J., Liu, L., Zhang, S., et al.: Imple- mentation and prospective clinical validation of ai-based planning and shimming techniques in cardiac mri. Medical Physics49(1), 129–143 (2022)

2022

Show all 29 references

[9]

European Radiology35, 3927–3936 (2025)

Glessgen, C., Crowe, L.A., Wetzl, J., Schmidt, M., Yoon, S.S., Vallée, J.P., Deux, J.F.: Automated vs manual cardiac mri planning: a single-center prospective eval- uation of reliability and scan times. European Radiology35, 3927–3936 (2025)

2025
[10]

In: IEEE Conference on Computer Vision and Pattern Recognition

He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: IEEE Conference on Computer Vision and Pattern Recognition. pp. 770–778 (2016)

2016
[11]

Journal of Cardiovascular Magnetic Resonance13(1), P39 (2011)

Kellman, P., Lu, X., Jolly, M.P., Bi, X., Kroeker, R., Schmidt, M., et al.: Au- tomatic lv localization and view planning for cardiac mri acquisition. Journal of Cardiovascular Magnetic Resonance13(1), P39 (2011)

2011
[12]

In: IEEE Conference on Computer Vision and Pattern Recognition

Lin, T.Y., Dollár, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: IEEE Conference on Computer Vision and Pattern Recognition. pp. 2117–2125 (2017)

2017
[13]

IEEE Transactions on Pattern Analysis and Machine Intelligence42(2), 318–327 (2020)

Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollár, P.: Focal loss for dense object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence42(2), 318–327 (2020)

2020
[14]

In: International Conference on Learning Representations (2018)

Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2018)

2018
[15]

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

Lu, X., Jolly, M.P., Georgescu, B., Hayes, C., Speier, P., Schmidt, M., et al.: Au- tomatic view planning for cardiac mri acquisition. In: International Conference on Medical Image Computing and Computer Assisted Intervention. pp. 479–486 (2011)

2011
[16]

IEEE Transactions on Pattern Analysis and Machine Intelligence46(5), 3537–3556 (2024)

Ma, X., Ouyang, W., Simonelli, A., Ricci, E.: 3d object detection from images for autonomous driving: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence46(5), 3537–3556 (2024)

2024
[17]

IEEE Journal of Biomedical and Health Informatics 27(7), 3302–3313 (2023)

Martín-Isla, C., Campello, V.M., Izquierdo, C., Kushibar, K., Sendra-Balcells, C., Gkontra, P., et al.: Deep learning segmentation of the right ventricle in cardiac mri: The m&ms challenge. IEEE Journal of Biomedical and Health Informatics 27(7), 3302–3313 (2023)

2023
[18]

In: Conference on Neural Information Processing Systems

Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., et al.: Py- torch: An imperative style, high-performance deep learning library. In: Conference on Neural Information Processing Systems. vol. 32, pp. 8024–8035 (2019)

2019
[19]

In: IEEE/CVF International Conference on Computer Vision

Tian, Z., Shen, C., Chen, H., He, T.: Fcos: Fully convolutional one-stage object detection. In: IEEE/CVF International Conference on Computer Vision. pp. 9626– 9635 (2019)

2019
[20]

Frontiers in Cardiovascular Medicine8(2021)

Vergani, V., Razavi, R., Puyol-Antón, E., Ruijsink, B.: Deep learning for classifica- tion and selection of cine cmr images to achieve fully automated quality-controlled cmr analysis from scanner to report. Frontiers in Cardiovascular Medicine8(2021)

2021
[21]

In: IEEE/CVF International Conference on Computer Vision Workshops

Wang, T., Zhu, X., Pang, J., Lin, D.: Fcos3d: Fully convolutional one-stage monoc- ular 3d object detection. In: IEEE/CVF International Conference on Computer Vision Workshops. pp. 913–922 (2021)

2021
[22]

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

Wei, D., Ma, K., Zheng, Y.: Training automatic view planner for cardiac mr imag- ing via self-supervision by spatial relationship between views. In: International Conference on Medical Image Computing and Computer Assisted Intervention. pp. 526–536 (2021)

2021
[23]

IEEE Trans- actions on Pattern Analysis and Machine Intelligence45(4), 4335–4354 (2023) LV Localization and SAX Plane Estimation from Arbitrary Slice 11

Yang, X., Zhang, G., Yang, X., Zhou, Y., Wang, W., Tang, J., et al.: Detecting rotated objects as gaussian distributions and its 3-d generalization. IEEE Trans- actions on Pattern Analysis and Machine Intelligence45(4), 4335–4354 (2023) LV Localization and SAX Plane Estimation...

2023
[24]

In: IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (2023)

Yu, Y., Da, F.: Phase-shifting coder: Predicting accurate orientation in oriented ob- ject detection. In: IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (2023)

2023
[25]

IEEE Transactions on Pattern Analysis and Machine Intelligence46(10), 6494–6508 (2024)

Yu, Y., Da, F.: On boundary discontinuity in angle regression based arbitrary oriented object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence46(10), 6494–6508 (2024)

2024
[26]

In: IEEE 23rd International Symposium on Biomedical Imaging (2026)

Yu, Y., Martin, P., Bu, Z., Liu, Y., Zheng, Y.Y., Simonetti, O., Han, Y., Xue, Y.: Uncertainty-aware structured data extraction from full cmr reports via distilled llms. In: IEEE 23rd International Symposium on Biomedical Imaging (2026)

2026
[27]

arXiv Preprint arXiv:2512.18176 (2025)

Zhang, Z., Yu, Y., Zhu, S., Aly, A., Gao, Y., Gu, N., Xue, Y.: Atlas is your per- fect context: One-shot customization for generalizable foundational medical image segmentation. arXiv Preprint arXiv:2512.18176 (2025)

2025 arXiv
[28]

In: European Conference on Computer Vision

Zheng, Y., Zhang, D., Xie, S., Lu, J., Zhou, J.: Rotation-robust intersection over union for 3d object detection. In: European Conference on Computer Vision. pp. 464–480 (2020)

2020
[29]

In: 30th ACM International Conference on Multimedia

Zhou, Y., Yang, X., Zhang, G., Wang, J., Liu, Y., Hou, L., et al.: Mmrotate: A rotated object detection benchmark using pytorch. In: 30th ACM International Conference on Multimedia. pp. 7331–7334 (2022)

2022

Pith tools

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

[1] [1]

European Heart Journal - Cardiovascular Imaging20(Supplement2), jez116.025 (2019)

Aung, N., Lee, A.M., Sanghvi, M.M., Fung, K., Paiva, J.M., Thomson, R.J., et al.: Automatic classification of cmr image sequences with convolutional neu- ral networks. European Heart Journal - Cardiovascular Imaging20(Supplement2), jez116.025 (2019)

2019

[2] [2]

In: Medical Imaging 2011: Image Process- ing

Bağci, U., Udupa, J.K., Chen, X.: Orientation estimation of anatomical structures in medical images for object recognition. In: Medical Imaging 2011: Image Process- ing. vol. 7962, p. 79622L. International Society for Optics and Photonics (2011)

2011

[3] [3]

Bernard, O., Lalande, A., Zotti, C., Cervenansky, F., Yang, X., Heng, P.A., et al.: Deep learning techniques for automatic mri cardiac multi-structures segmenta- tion and diagnosis: Is the problem solved? IEEE Transactions on Medical Imaging 37(11), 2514–2525 (2018)

2018

[4] [4]

Radiology: Artificial Intelligence1(6), e180069 (2019)

Blansit, K., Retson, T., Masutani, E., Bahrami, N., Hsiao, A.: Deep learning–based prescription of cardiac mri planes. Radiology: Artificial Intelligence1(6), e180069 (2019)

2019

[5] [5]

Journal of Magnetic Resonance Imaging63(3), 891–903 (2026)

Böttcher, B., Meinel, F.G., Deyerberg, K.K., Watzke, L.M., Manzke, M., Gorodezky,M.,etal.:Fullyautomatedplaneprescriptionincardiacmri:Aprospec- tive cohort study. Journal of Magnetic Resonance Imaging63(3), 891–903 (2026)

2026

[6] [6]

IEEE Transactions on Medical Imaging40(12), 3543–3554 (2021)

Campello, V.M., Gkontra, P., Izquierdo, C., Martín-Isla, C., Sojoudi, A., Full, P.M., et al.: Multi-centre, multi-vendor and multi-disease cardiac segmentation: The m&ms challenge. IEEE Transactions on Medical Imaging40(12), 3543–3554 (2021)

2021

[7] [7]

Clinical Imaging82, 121–126 (2022) 10 Y

Chauhan, D., Anyanwu, E., Goes, J., Besser, S.A., Anand, S., Madduri, R., et al.: Comparison of machine learning and deep learning for view identification from cardiac magnetic resonance images. Clinical Imaging82, 121–126 (2022) 10 Y. Yu et al

2022

[8] [8]

Medical Physics49(1), 129–143 (2022)

Edalati, M., Zheng, Y., Watkins, M.P., Chen, J., Liu, L., Zhang, S., et al.: Imple- mentation and prospective clinical validation of ai-based planning and shimming techniques in cardiac mri. Medical Physics49(1), 129–143 (2022)

2022