Pith. sign in

REVIEW 3 major objections 5 minor 41 references

1.13M-param model outdoes 3D deep nets on heart echo, no backprop

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 07:31 UTC pith:UGD624JB

load-bearing objection The multi-task GL pipeline is a real engineering contribution, but the classification accuracy is likely inflated by training the encoder on masks, so treat the headline 94.3% with caution. the 3 major comments →

arxiv 2601.19743 v3 pith:UGD624JB submitted 2026-01-27 eess.IV cs.CVcs.LG

Interpretable and backpropagation-free Green Learning for efficient multi-task echocardiographic segmentation and classification

Jyun-Ping Kao , Jiaxin Yang , C.-C. Jay Kuo , Jonghye Woo This is my paper
classification eess.IV cs.CVcs.LG
keywords green learningechocardiographyLVEF classificationleft ventricle segmentationVoxelHopSaab transformmulti-task learningbackpropagation-free
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper sets out to prove that a feed-forward, backpropagation-free pipeline—feature extraction by PCA-based multi-scale transforms, segmentation by coarse-to-fine residual regression, classification by a gradient-boosted decision tree—can handle both left-ventricle segmentation and ejection-fraction classification from echocardiograms at the level of, or better than, large 3D deep networks. On the EchoNet-Dynamic dataset, the proposed multi-task model reaches 94.3% three-class accuracy and a Dice score of 0.912, edging out several advanced 3D deep-learning baselines while using roughly one tenth to one fortieth of their parameters. A sympathetic reader would care because, if correct, it means clinical cardiac image analysis does not need millions of parameters or backpropagation; small, transparent models can do the job, with interpretable features and a greatly reduced energy footprint.

Core claim

The central claim is that a multi-task Green Learning (MTGL) framework—combining an unsupervised VoxelHop encoder (four cascaded PCA/Saab units), a four-level coarse-to-fine residual regression decoder, and a gradient-boosted decision-tree classifier on pooled multi-scale features—performs simultaneous left-ventricle segmentation and ejection-fraction classification on par with or better than advanced 3D deep-learning models, while using only about 1.13 million parameters. The authors further demonstrate that the learned linear filters are interpretable: the shallowest hop yields edge-like responses that localize the endocardial border and papillary muscles, while the deepest hop behaves lik

What carries the argument

The VoxelHop encoder: four cascaded units that use the Saab transform—a channel-wise PCA with a DC/AC split, where the DC component captures local mean intensity and the AC components capture decorrelated directional variations—to build hierarchical spatio-temporal features without any backpropagation. The encoder is shared by both tasks: the segmentation decoder routes its multi-scale features through four resolution levels of gradient-boosted residual regressors that progressively correct the predicted mask toward the full resolution, and the classification decoder pools the same features into a global descriptor that a gradient-boosted decision tree maps to three ejection-fraction categor

Load-bearing premise

The classification performance depends on the VoxelHop encoder having been trained on ground-truth mask volumes; if the encoder were trained on images alone, the reported 94.3% accuracy might not hold.

What would settle it

Retrain the VoxelHop encoder using only image volumes (no masks) and rerun the classification decoder on the resulting features; if accuracy drops materially below 94.3%, mask-derived information was essential to the reported result. Additionally, run the full pipeline on an external multi-vendor echo dataset; a large accuracy drop would indicate that the result is specific to the EchoNet-Dynamic acquisition protocol.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Echocardiogram analysis can run without backpropagation: if the reported results reproduce, gradient-free PCA-based encoders plus tree decoders are a viable alternative to deep networks for segmentation and classification.
  • The model's parameter count (1.13 million) and energy consumption (about 5 kWh for training) are an order of magnitude lower than the compared deep networks, which would ease deployment on portable ultrasound devices and in energy-constrained settings.
  • The feature-space interpretation gives clinicians a visual chain from image to decision: shallow hops trace borders, deep hops track global longitudinal motion, so errors can be attributed to a specific stage rather than buried in a black box.
  • The ablation study shows that combining all four hops yields the best accuracy (94.3%), and that the deepest hop alone captures most of the class-relevant variation, suggesting that multi-scale information is complementary for ejection-fraction grouping.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper states in Section 2.1.3 that the VoxelHop encoder was trained on image and mask volumes, so the 'unsupervised' representation used for classification already contains segmentation-label information. A natural test of the label-free claim would be to retrain the encoder on images alone and measure the accuracy change; the paper does not report this experiment.
  • The 94.3% figure comes from a single public dataset, and the borderline 40–50% ejection-fraction class was oversampled to reach it. Whether the method transfers to other vendors, image qualities, and patient mixes is untested; a multi-center evaluation would be the decisive check.
  • Because the deepest hop's response approximates a temporal derivative of through-plane motion, the classifier could latch onto annular descent and other proxy motions rather than the ventricular cavity itself. This suggests a falsifiable probe: curated cases with isolated wall-motion abnormalities but preserved ejection fraction should cause the classifier to misclassify if it relies on motion pro
  • The gradient-boosted tree decoder is itself an opaque component; the paper's interpretability claims apply to the encoder, not to the decision-tree mapping from pooled features to class. A future extension could replace the tree with a linear or rule-based head to maintain the white-box property end to end.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes MTGL, a multi-task feed-forward 'green learning' framework that performs simultaneous left-ventricle (LV) segmentation and LVEF classification on the EchoNet-Dynamic dataset. A VoxelHop encoder built from PCA/Saab transforms extracts hierarchical spatio-temporal features; a coarse-to-fine XGBoost regression decoder produces the segmentation mask; a separate XGBoost classifier on pooled encoder features predicts three LVEF classes. The authors report DSC 0.912, IoU 0.838, classification accuracy 94.3%, balanced accuracy 80.8%, and 1.13M parameters, claiming to outperform 3D V-Net, 3D UNETR, 3D U-Net, and 3D nnU-Net baselines trained under the same protocol.

Significance. If the classification result survives an image-only, non-oracle evaluation, the paper would be a valuable demonstration that backpropagation-free, interpretable models can compete with 3D deep networks on echocardiography while using far fewer parameters. The controlled 3D baselines, per-hop energy analysis, and ablation study are useful contributions. However, the central classification claim is jeopardized by the explicit statement that the VoxelHop encoder was trained on image and mask volumes, which leaks segmentation ground truth into the classification feature vector. The possible use of oracle EDV/ESV frames is a further threat to the 'state-of-the-art' claim. The significance is therefore conditional on the results of the requested ablations and on a precise statement of the input representation.

major comments (3)
  1. [§2.1.3] The classification pipeline is not label-free. The text states that 'the VoxelHop encoder was trained on image and mask volumes' and that the pooled descriptor H 'inherently incorporate[s] knowledge of the LV region and boundaries.' This directly contradicts the 'unsupervised' claim in §2.1.1. Because the XGBoost classifier operates on H, the reported 0.9429 accuracy in Table 1 can reflect mask-derived information rather than a general green-learning capability. Please retrain the encoder on image volumes only and report classification accuracy and balanced accuracy; if masks are needed for the segmentation decoder, show an ablation that isolates their effect on the classification feature vector.
  2. [§2.1 (Data and Preprocessing)] The input description is ambiguous and potentially oracle-based. The text says, 'we stack the resized End-Diastolic Volume (EDV) and End-Systolic Volume (ESV) images from the same video series ... as a single dataset,' which suggests that the EDV/ESV frames are selected using ground-truth labels. If so, the LVEF classifier is given the exact frames of maximal and minimal volume, substantially simplifying the task and making a comparison to EchoNet-Dynamic's video-based model inappropriate. Please clarify the exact input tensor (dimensions, number of channels, whether the 12-frame sequence is independent of the EDV/ESV labels) and either evaluate on non-oracle frames or limit the claim to 'matched 3D baselines with oracle phase inputs.'
  3. [Table 2 and §3.3] The reported 1.13M parameter count for 'Ours' is not supported by a breakdown. The segmentation decoder uses multiple XGBoost regressors at four resolution levels and the classification decoder uses an additional XGBoost classifier; these tree ensembles have leaf weights and split thresholds that are trained parameters. If these are excluded from the count, the 'order of magnitude fewer parameters' claim is misleading. Please report the full parameter budget, including all XGBoost trees, and state the tree counts, depths, and other hyperparameters used. This is necessary to fairly compare model complexity with the DL baselines.
minor comments (5)
  1. [Table 5] The row label 'Class 2 & 2' likely should read 'Class 2 & 3'. Also specify the oversampling procedure (e.g., duplication vs. synthetic augmentation) and confirm it was applied only to the training set.
  2. [Eq. (2)] The formula is called GAP but includes both mean and max operations. This is not standard global average pooling; please rename it (e.g., 'mean-max pooling') or correct the description for consistency.
  3. [§3.1 and Abstract] The term 'state-of-the-art' is used in the abstract and results, but the comparison is limited to the four 3D baselines implemented here. The paper does not compare with EchoNet-Dynamic's original 2D video model or with other published 2D echocardiography results. Please either add such a comparison or rephrase to 'outperforms the matched 3D baselines in our controlled setting.'
  4. [§2.1] The input construction is confusing: 'size 112×112×2' vs. '12 consecutive frames ... concatenate their two-channel frames along the temporal axis.' Please provide the final tensor dimensions and clarify the relationship between the 12-frame sequence and the EDV/ESV images.
  5. [Overall] Reproducibility would be improved by releasing code and the full hyperparameter settings (XGBoost tree counts, depth, learning rate, PCA energy thresholds, oversampling multiplier, and any preprocessing normalization). The current manuscript leaves many of these unspecified.

Circularity Check

1 steps flagged

Classification accuracy is not label-free: §2.1.3 admits the VoxelHop encoder was trained on image and mask volumes, while EF classes are computed from the same expert tracings.

specific steps
  1. fitted input called prediction [Section 2.1.3, Classification decoder, final paragraph]
    "Furthermore, because the segmentation mask information was embedded into the feature extraction process (recall that the VoxelHop encoder was trained on image and mask volumes), the features H inherently incorporate knowledge of the LV region and boundaries. This effectively guides the classifier to focus on the relevant cardiac structures for predicting LVEF."

    The EF classes are derived from the expert tracings (Section 2.1: 'the patient's LVEF is discretized into three classes based on clinical cut-offs derived from the expert-verified manual tracings'), and the same tracings are converted into binary masks used to fit the VoxelHop encoder. Thus H is produced by PCA filters that were estimated on image+mask volumes, so the representation carries mask-derived knowledge of LV cavity geometry and boundary; EF is by definition a function of that geometry (EDV/ESV from the tracings). The XGBoost 'prediction' of EF category is therefore not an independent unsupervised classification but a readout of a representation fitted to the very labels from which the target is computed.

full rationale

The segmentation half of MTGL is a standard supervised coarse-to-fine regressor trained against binary masks; no circularity is apparent there. The circularity is in the classification half. The paper's own Section 2.1.3 explicitly says the VoxelHop encoder was trained on image and mask volumes and that H 'inherently incorporate[s] knowledge of the LV region and boundaries,' while Section 2.1 defines the three EF classes from expert-verified manual tracings—the same tracings that are converted to the training masks. Consequently, the 94.3% classification accuracy is not evidence of a label-free green-learning representation; it is a classifier fitted to features whose construction was supervised by the anatomical labels that determine EF. This does not make the whole paper circular: the segmentation metrics (DSC 0.912) stand on their own, and the PCA/energy/ablation analyses are informative about the method's behavior. But the headline multi-task claim of 'unsupervised VoxelHop encoder' plus superior EF classification is partially circular by construction because mask information is baked into H. A mask-free encoder ablation would be needed to support the advertised classification claim. This is a partial-circularity finding rather than a wholesale reduction of the paper.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The central results rest on standard PCA/tree-ensemble machinery plus several domain assumptions about EchoNet-Dynamic labels, the sufficiency of linear features, and the input-volume construction. No new physical or model-level entities are introduced beyond the existing VoxelHop/GUSL components.

free parameters (4)
  • Per-hop cumulative-energy threshold (K99 + safety margin) = Hop1: 26/26; Hop2: 185/702; Hop3: 378/5252; Hop4: 643/11102
    Section 3.6/Table 6: hand-set energy retention strategy inherited from PixelHop/VoxelHop; controls encoder width and is chosen from data spectra rather than task labels.
  • Oversampling multiplier for borderline LVEF class (Class 2) = 60 -> 480 training samples
    Section 3.5/Table 5: augmenting Class 2 to equal Class 1 count raises test accuracy from 0.8952 to 0.9429; selected after observing imbalance, so this is a post-hoc training-set choice.
  • SPP pooling grid (2x2x1) and GAP for middle levels = 2x2x1 SPP for F1/F4; GAP for F2/F3
    Section 2.1.3: pooling configuration is chosen by hand; changes descriptor dimensionality and classification accuracy.
  • XGBoost hyperparameters (tree count, depth, learning rate) = not reported
    XGBoost regressors/classifier throughout; no hyperparameters or search procedure given, so these are implicit fitted choices needed for reproduction.
axioms (5)
  • domain assumption EchoNet-Dynamic expert tracings and EF labels are reliable ground truth.
    All segmentation and classification metrics are computed against these labels; no independent adjudication is described.
  • standard math Saab/PCA produces a decorrelated orthogonal decomposition so that AC eigen-energies correspond to retained variance.
    Section 2.1.1 invokes the DC/AC split and PCA eigenvalues; this standard linear-algebra fact underlies the energy-based filter selection.
  • domain assumption Retaining 99% cumulative PCA energy per VoxelHop hop preserves task-relevant information.
    Section 3.6 justifies filter counts by eigen-energy; orthogonality bounds reconstruction error at 1%, but this does not guarantee that discarded ACs are irrelevant to EF classes or boundaries.
  • domain assumption Linear Saab/PCA features are sufficient to support both segmentation regression and LVEF classification.
    The entire framework uses linear transforms plus tree ensembles; the authors acknowledge linearity limits nonlinear dynamics (Discussion), yet the performance claim depends on this sufficiency.
  • domain assumption The input volume (12 consecutive frames or 2-channel EDV/ESV stack) captures the spatiotemporal contraction signal for LVEF.
    Section 2.1 describes both 2-channel and 12-frame constructions without clarifying which is used; the classifier's accuracy relies on the chosen input preserving ED-to-ES dynamics.

pith-pipeline@v1.3.0-alltime-deepseek · 16240 in / 16454 out tokens · 178223 ms · 2026-08-03T07:31:44.654240+00:00 · methodology

0 comments
read the original abstract

Echocardiography is a cornerstone for managing heart failure (HF), with Left Ventricular Ejection Fraction (LVEF) being a critical metric for guiding therapy. However, manual LVEF assessment suffers from high inter-observer variability, while existing Deep Learning (DL) models are often computationally intensive and data-hungry "black boxes" that impede clinical trust and adoption. Here, we propose a backpropagation-free multi-task Green Learning (MTGL) framework that performs simultaneous Left Ventricle (LV) segmentation and LVEF classification. Our framework integrates an unsupervised VoxelHop encoder for hierarchical spatio-temporal feature extraction with a multi-level regression decoder and an XG-Boost classifier. On the EchoNet-Dynamic dataset, our MTGL model achieves state-of-the-art classification and segmentation performance, attaining a classification accuracy of 94.3% and a Dice Similarity Coefficient (DSC) of 0.912, significantly outperforming several advanced 3D DL models. Crucially, our model achieves this with over an order of magnitude fewer parameters, demonstrating exceptional computational efficiency. This work demonstrates that the GL paradigm can deliver highly accurate, efficient, and interpretable solutions for complex medical image analysis, paving the way for more sustainable and trustworthy artificial intelligence in clinical practice.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 2 linked inside Pith

  1. [1]

    Nature Reviews Cardiology17(9), 559–573 (2020)

    Borlaug, B.A.: Evaluation and management of heart failure with preserved ejection fraction. Nature Reviews Cardiology17(9), 559–573 (2020)

    2020

  2. [2]

    Nature Reviews Cardiology13(6), 368–378 (2016)

    Ziaeian, B., Fonarow, G.C.: Epidemiology and aetiology of heart failure. Nature Reviews Cardiology13(6), 368–378 (2016)

    2016

  3. [3]

    Circulation research119(2), 357–374 (2016)

    Omar, A.M.S., Bansal, M., Sengupta, P.P.: Advances in echocardiographic imag- ing in heart failure with reduced and preserved ejection fraction. Circulation research119(2), 357–374 (2016)

    2016

  4. [4]

    Journal of Nuclear Medicine56(Supplement 4), 31–38 (2015)

    Marwick, T.H.: The role of echocardiography in heart failure. Journal of Nuclear Medicine56(Supplement 4), 31–38 (2015)

    2015

  5. [5]

    Journal of the American College of Cardiology79(17), 263–421 (2022)

    Heidenreich, P.A., Bozkurt, B., Aguilar, D., Allen, L.A., Byun, J.J., Colvin, M.M., Deswal, A., Drazner, M.H., Dunlay, S.M., Evers, L.R.,et al.: 2022 aha/ac- c/hfsa guideline for the management of heart failure: a report of the american college of cardiology/american heart association joint committee on clinical prac- tice guidelines. Journal of the Ameri...

    2022

  6. [6]

    Journal of the American College of Cardiology83(15), 1444–1488 (2024)

    Maddox, T.M., Januzzi Jr, J.L., Allen, L.A., Breathett, K., Brouse, S., Butler, J., Davis, L.L., Fonarow, G.C., Ibrahim, N.E., Lindenfeld, J.,et al.: 2024 acc expert consensus decision pathway for treatment of heart failure with reduced ejection fraction: a report of the american college of cardiology solution set oversight committee. Journal of the Ameri...

    2024

  7. [7]

    Jama329(10), 827–838 (2023)

    Redfield, M.M., Borlaug, B.A.: Heart failure with preserved ejection fraction: a review. Jama329(10), 827–838 (2023)

    2023

  8. [8]

    The international journal of cardiovascular imaging31(7), 1303–1314 (2015) 23

    Cole, G.D., Dhutia, N.M., Shun-Shin, M.J., Willson, K., Harrison, J., Raphael, C.E., Zolgharni, M., Mayet, J., Francis, D.P.: Defining the real-world repro- ducibility of visual grading of left ventricular function and visual estimation of left ventricular ejection fraction: impact of image quality, experience and accred- itation. The international journa...

    2015

  9. [9]

    The Lancet357(9262), 1107–1117 (2001)

    Glikson, M., Friedman, P.A.: The implantable cardioverter defibrillator. The Lancet357(9262), 1107–1117 (2001)

    2001

  10. [10]

    European Journal of Heart Failure23(4), 578–589 (2021)

    Turgeon, R.D., Barry, A.R., Hawkins, N.M., Ellis, U.M.: Pharmacotherapy for heart failure with reduced ejection fraction and health-related quality of life: a systematic review and meta-analysis. European Journal of Heart Failure23(4), 578–589 (2021)

    2021

  11. [11]

    Journal of echocardiography17(1), 10–16 (2019)

    Klaeboe, L.G., Edvardsen, T.: Echocardiographic assessment of left ventricular systolic function. Journal of echocardiography17(1), 10–16 (2019)

    2019

  12. [12]

    Journal of the American College of Cardiology66(13), 1456–1466 (2015)

    Knackstedt, C., Bekkers, S.C., Schummers, G., Schreckenberg, M., Muraru, D., Badano, L.P., Franke, A., Bavishi, C., Omar, A.M.S., Sengupta, P.P.: Fully automated versus standard tracking of left ventricular ejection fraction and lon- gitudinal strain: the fast-efs multicenter study. Journal of the American College of Cardiology66(13), 1456–1466 (2015)

    2015

  13. [13]

    The international journal of cardiovascular imaging26(1), 57–64 (2010)

    Luijnenburg, S.E., Robbers-Visser, D., Moelker, A., Vliegen, H.W., Mulder, B.J., Helbing, W.A.: Intra-observer and interobserver variability of biventricular func- tion, volumes and mass in patients with congenital heart disease measured by cmr imaging. The international journal of cardiovascular imaging26(1), 57–64 (2010)

    2010

  14. [14]

    Echocardiography38(4), 582– 589 (2021)

    Kouris, N.T., Kostopoulos, V.S., Psarrou, G.A., Kostakou, P.M., Tzavara, C., Olympios, C.D.: Left ventricular ejection fraction and global longitudinal strain variability between methodology and experience. Echocardiography38(4), 582– 589 (2021)

    2021

  15. [15]

    Journal of the American Heart Association13(8), 032257 (2024)

    Christersson, M., Gustafsson, S., Lampa, E., Almstedt, M., Cars, T., Bodeg ˚ ard, J., Arefalk, G., Sundstr¨ om, J.: Usefulness of heart failure categories based on left ventricular ejection fraction. Journal of the American Heart Association13(8), 032257 (2024)

    2024

  16. [16]

    In: International Conference on Medical Image Computing and Computer-assisted Intervention, pp

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: International Conference on Medical Image Computing and Computer-assisted Intervention, pp. 234–241 (2015). Springer

    2015

  17. [17]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

    Azad, R., Aghdam, E.K., Rauland, A., Jia, Y., Avval, A.H., Bozorgpour, A., Karimijafarbigloo, S., Cohen, J.P., Adeli, E., Merhof, D.: Medical image segmen- tation review: The success of u-net. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

    2024

  18. [18]

    Nature580(7802), 252–256 (2020)

    Ouyang, D., He, B., Ghorbani, A., Yuan, N., Ebinger, J., Langlotz, C.P., Heiden- reich, P.A., Harrington, R.A., Liang, D.H., Ashley, E.A.,et al.: Video-based ai for beat-to-beat assessment of cardiac function. Nature580(7802), 252–256 (2020)

    2020

  19. [19]

    Journal of big Data8(1), 53 (2021)

    Alzubaidi, L., Zhang, J., Humaidi, A.J., Al-Dujaili, A., Duan, Y., Al-Shamma, O., Santamar ´ ıa, J., Fadhel, M.A., Al-Amidie, M., Farhan, L.: Review of deep 24 learning: concepts, cnn architectures, challenges, applications, future directions. Journal of big Data8(1), 53 (2021)

    2021

  20. [20]

    Journal of Cardiovascular Imaging29(3), 193 (2021)

    Yoon, Y.E., Kim, S., Chang, H.-J.: Artificial intelligence and echocardiography. Journal of Cardiovascular Imaging29(3), 193 (2021)

    2021

  21. [21]

    Journal of Visual Communication and Image Representation41, 406–413 (2016)

    Kuo, C.-C.J.: Understanding convolutional neural networks with a mathematical model. Journal of Visual Communication and Image Representation41, 406–413 (2016)

    2016

  22. [22]

    Journal of Visual Communication and Image Representation90, 103685 (2023)

    Kuo, C.-C.J., Madni, A.M.: Green learning: Introduction, examples and outlook. Journal of Visual Communication and Image Representation90, 103685 (2023)

    2023

  23. [23]

    NPJ digital medicine5(1), 156 (2022)

    Chen, H., Gomez, C., Huang, C.-M., Unberath, M.: Explainable medical imag- ing ai needs human-centered design: guidelines and evidence from a systematic review. NPJ digital medicine5(1), 156 (2022)

    2022

  24. [24]

    Journal of Visual Communication and Image Representation 70, 102749 (2020)

    Chen, Y., Kuo, C.-C.J.: Pixelhop: A successive subspace learning (ssl) method for object recognition. Journal of Visual Communication and Image Representation 70, 102749 (2020)

    2020

  25. [25]

    IEEE Transactions on Multimedia22(7), 1744–1755 (2020)

    Zhang, M., You, H., Kadam, P., Liu, S., Kuo, C.-C.J.: Pointhop: An explainable machine learning method for point cloud classification. IEEE Transactions on Multimedia22(7), 1744–1755 (2020)

    2020

  26. [26]

    IEEE journal of biomedical and health informatics26(3), 1128– 1139 (2021)

    Liu, X., Xing, F., Yang, C., Kuo, C.-C.J., Babu, S., El Fakhri, G., Jenkins, T., Woo, J.: Voxelhop: Successive subspace learning for als disease classification using structural mri. IEEE journal of biomedical and health informatics26(3), 1128– 1139 (2021)

    2021

  27. [27]

    arXiv preprint arXiv:2506.23688 (2025)

    Yang, J., Magoulianitis, V., Alexander, C.A.C., Xue, J., Kaneko, M., Cacciamani, G., Abreu, A., Duddalwar, V., Kuo, C.-C.J., Gill, I.S., et al.: Gusl: A novel and efficient machine learning model for prostate segmentation on mri. arXiv preprint arXiv:2506.23688 (2025)

    Pith/arXiv arXiv 2025

  28. [28]

    JAMA cardiology6(6), 624–632 (2021)

    Narang, A., Bae, R., Hong, H., Thomas, Y., Surette, S., Cadieu, C., Chaudhry, A., Martin, R.P., McCarthy, P.M., Rubenson, D.S.,et al.: Utility of a deep-learning algorithm to guide novices to acquire echocardiograms for limited diagnostic use. JAMA cardiology6(6), 624–632 (2021)

    2021

  29. [29]

    BMC medical informatics and decision making20(1), 310 (2020)

    Amann, J., Blasimme, A., Vayena, E., Frey, D., Madai, V.I., Consortium, P.: Explainability for artificial intelligence in healthcare: a multidisciplinary perspective. BMC medical informatics and decision making20(1), 310 (2020)

    2020

  30. [30]

    In: Proceedings of the 22nd Acm Sigkdd International Conference on Knowledge Discovery and Data Mining, pp

    Chen, T., Guestrin, C.: Xgboost: A scalable tree boosting system. In: Proceedings of the 22nd Acm Sigkdd International Conference on Knowledge Discovery and Data Mining, pp. 785–794 (2016) 25

    2016

  31. [31]

    Wiley interdisciplinary reviews: computational statistics2(4), 433–459 (2010)

    Abdi, H., Williams, L.J.: Principal component analysis. Wiley interdisciplinary reviews: computational statistics2(4), 433–459 (2010)

    2010

  32. [32]

    American heart journal146(3), 388–397 (2003)

    McGowan, J.H., Cleland, J.G.: Reliability of reporting left ventricular systolic function by echocardiography: a systematic review of 3 methods. American heart journal146(3), 388–397 (2003)

    2003

  33. [33]

    IEEE transactions on pattern analysis and machine intelligence37(9), 1904–1916 (2015)

    He, K., Zhang, X., Ren, S., Sun, J.: Spatial pyramid pooling in deep convolu- tional networks for visual recognition. IEEE transactions on pattern analysis and machine intelligence37(9), 1904–1916 (2015)

    1904

  34. [34]

    arXiv preprint arXiv:1312.4400 (2013)

    Lin, M., Chen, Q., Yan, S.: Network in network. arXiv preprint arXiv:1312.4400 (2013)

    Pith/arXiv arXiv 2013

  35. [35]

    In: 2016 Fourth International Conference on 3D Vision (3DV), pp

    Milletari, F., Navab, N., Ahmadi, S.-A.: V-net: Fully convolutional neural net- works for volumetric medical image segmentation. In: 2016 Fourth International Conference on 3D Vision (3DV), pp. 565–571 (2016). Ieee

    2016

  36. [36]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp

    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: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp. 574–584 (2022)

    2022

  37. [37]

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

    Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods18(2), 203–211 (2021)

    2021

  38. [38]

    Clinical Physiology and Functional Imaging41(5), 443–451 (2021)

    Støylen, A., Daae, A.S.: Physiological significance of pre-and post-ejection left ventricular tissue velocities and relations to mitral and aortic valve closures. Clinical Physiology and Functional Imaging41(5), 443–451 (2021)

    2021

  39. [39]

    data from the hunt study

    Støylen, A., Mølmen, H.E., Dalen, H.: Regional motion of the av-plane is related to the cardiac anatomy and deformation of the av-plane. data from the hunt study. Clinical physiology and functional imaging43(6), 453–462 (2023)

    2023

  40. [40]

    Quinn, T.P., Jacobs, S., Senadeera, M., Le, V., Coghlan, S.: The three ghosts of medical ai: Can the black-box present deliver? Artificial intelligence in medicine 124, 102158 (2022)

    2022

  41. [41]

    npj Digital Medicine8(1), 341 (2025) 26

    Chao, C.-J., Gu, Y.R., Kumar, W., Xiang, T., Appari, L., Wu, J., Farina, J.M., Wraith, R., Jeong, J., Arsanjani, R.,et al.: Foundation versus domain- specific models for left ventricular segmentation on cardiac ultrasound. npj Digital Medicine8(1), 341 (2025) 26

    2025