arxiv: 2604.03313 · v1 · submitted 2026-03-31 · 💻 cs.CV

Recognition: unknown

CardioSAM: Topology-Aware Decoder Design for High-Precision Cardiac MRI Segmentation

Ujjwal Jain

Authors on Pith no claims yet

Pith reviewed 2026-05-08 02:24 UTC · model gemini-3-flash-preview

classification 💻 cs.CV

keywords Cardiac MRISegmentationFoundation ModelsSAMTopology-AwareMedical ImagingACDC Benchmark

0 comments

The pith

A topology-aware decoder allows general-purpose vision models to segment cardiac MRIs with higher precision than human experts.

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

Cardiac MRI analysis is critical for diagnosing heart disease, but manual segmentation of heart chambers is slow and varies between doctors. This paper proposes CardioSAM, a hybrid model that uses a powerful general-purpose vision encoder to see the image and a new, specialized decoder to interpret it. The decoder is designed with modules that understand the specific shapes and boundaries of the heart, ensuring the results are medically plausible. By combining general visual intelligence with specific anatomical rules, the system reaches a level of accuracy that matches or beats human specialists.

Core claim

The author demonstrates that the Segment Anything Model (SAM), despite being trained on natural images, can be adapted for high-precision cardiac segmentation by replacing its decoder with one that understands heart geometry. The resulting model, CardioSAM, achieves a Dice score of 93.39%, which is nearly 4% higher than standard medical imaging baselines and exceeds the typical level of agreement between human experts. This indicates that the primary barrier to using foundation models in medicine is not the encoder's lack of medical knowledge, but the decoder's lack of anatomical constraints.

What carries the argument

The Cardiac-Specific Attention module, which integrates anatomical topological priors into the decoding process to ensure heart structures maintain their biologically correct spatial relationships and shapes.

If this is right

Automated cardiac measurements, such as ejection fraction, could reach clinical consistency levels that rival or exceed human experts.
The time required for cardiac MRI analysis could be reduced from minutes of manual labor to seconds of automated processing.
Foundation models trained on natural images can be effectively repurposed for specialized medical tasks without needing massive new datasets or retraining.
Boundary-sensitive refinement modules could be applied to other medical imaging tasks where precise tissue interfaces are critical, such as neuroimaging.

Where Pith is reading between the lines

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

The success of a frozen encoder suggests that high-level visual concepts like edges and textures are universal enough that medical-specific encoders may eventually become obsolete.
This architecture suggests that re-decoding general features through a domain-aware lens is more effective than standard finetuning for high-precision tasks.
If topological priors are the missing link for foundation models, we may see a shift toward geometric deep learning in medical imaging rather than simply increasing model size.

Load-bearing premise

The model assumes that a vision system trained on everyday photographs provides a feature set that is sufficient to capture the subtle, low-contrast boundaries of cardiac tissues in MRI scans.

What would settle it

The claim would be falsified if the model's accuracy drops significantly below standard baselines when applied to hearts with severe structural abnormalities or surgical implants that deviate from the topological priors used in the decoder.

Figures

Figures reproduced from arXiv: 2604.03313 by Ujjwal Jain.

**Figure 1.** Figure 1: Class Distributions II. PROPOSED METHODOLOGY A. Formal Problem Definition and Topological Overview The fundamental objective of cardiovascular magnetic resonance (CMR) image segmentation can be rigorously formalized as discovering an optimal, highly non-linear mapping function Φ : X → Y. Here, X ⊂ R B×1×H×W represents the highdimensional manifold of input grayscale CMR slices (with batch size B, spatial h… view at source ↗

**Figure 2.** Figure 2: CardioSAM architecture diagram with all modules and data flow view at source ↗

**Figure 3.** Figure 3: Cardiac MRI medical images (ACDC dataset). (a) Normal Cardiacs; (b) Previous Myocardial Infarction; (c) Hypertrophic view at source ↗

**Figure 4.** Figure 4: Learning Rate Sensitivity TABLE II: CardioSAM Memory Usage Breakdown Component Memory (GB) Optimization SAM Encoder 3.6 Frozen weights Cardiac Decoder 1.8 Efficient design Input/Output Buffers 0.6 Minimal overhead Intermediate Features 0.3 Memory reuse Total 6.3 Optimized IV. RESULTS AND DISCUSSIONS This section presents the experimental results and a detailed analysis of CardioSAM’s performance. In Subsec… view at source ↗

**Figure 5.** Figure 5: CardioSAM Segmentation Results on ACDC Test Set view at source ↗

**Figure 6.** Figure 6: Segmentation Results based on probability maps view at source ↗

read the original abstract

Accurate segmentation of cardiac structures in cardiovascular magnetic resonance (CMR) images is essential for reliable diagnosis and treatment of cardiovascular diseases. However, manual segmentation remains time-consuming and suffers from significant inter-observer variability. Recent advances in deep learning, particularly foundation models such as the Segment Anything Model (SAM), demonstrate strong generalization but often lack the boundary precision required for clinical applications. To address this limitation, we propose CardioSAM, a hybrid architecture that combines the generalized feature extraction capability of a frozen SAM encoder with a lightweight, trainable cardiac-specific decoder. The proposed decoder introduces two key innovations: a Cardiac-Specific Attention module that incorporates anatomical topological priors, and a Boundary Refinement Module designed to improve tissue interface delineation. Experimental evaluation on the ACDC benchmark demonstrates that CardioSAM achieves a Dice coefficient of 93.39%, IoU of 87.61%, pixel accuracy of 99.20%, and HD95 of 4.2 mm. The proposed method surpasses strong baselines such as nnU-Net by +3.89% Dice and exceeds reported inter-expert agreement levels (91.2%), indicating its potential for reliable and clinically applicable cardiac segmentation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

CardioSAM achieves high-precision cardiac segmentation by combining a frozen SAM encoder with a custom decoder that uses spatial priors and boundary-weighted loss.

read the letter

The punchline here is that CardioSAM hits 93.39% Dice on the ACDC benchmark, which puts it at the very top of current performance. It achieves this by keeping the massive SAM ViT encoder frozen and only training a lightweight decoder. If you are looking for a clear recipe for adapting foundation models to specific clinical tasks without massive compute, this paper is a solid reference.

What the paper does well is focus on the 'blobby' output problem inherent in SAM. The Boundary Refinement Module and the associated loss function are well-executed. They specifically target tissue interfaces, which is where clinical segmentation usually fails. The results are transparently reported, and the modular design of the decoder is easy to understand and potentially easy to port to other anatomical structures.

There are two soft spots you should note. First, the 'topology-aware' claim is a bit of a stretch. The Cardiac-Specific Attention module basically uses a spatial template—an average of the training masks—to guide the attention. It is a spatial location prior, not a mathematical topological constraint. This works wonderfully on the ACDC dataset because the images are highly standardized and centered, but it might not be robust to the anatomical variance or off-center framing you see in messy clinical environments.

Second, the comparison to nnU-Net is likely tilted in the author's favor. They report an nnU-Net baseline of 89.5% Dice, whereas most well-tuned nnU-Net implementations on ACDC hit 91.5% or higher. The actual margin of improvement is probably closer to 1-2% rather than the 3.9% claimed.

Even with those caveats, the paper is a serious effort. It proves that you don't need to fine-tune the backbone to get state-of-the-art results if your decoder is designed with the right priors. This is useful for anyone working on medical FM adaptation. It definitely deserves a serious referee.

Referee Report

3 major / 3 minor

Summary. The manuscript presents CardioSAM, a segmentation framework for cardiac MRI that leverages a frozen Segment Anything Model (SAM) image encoder coupled with a novel decoder. The decoder includes two primary components: a Cardiac-Specific Attention (CSA) module, which utilizes an anatomical template derived from training data to provide spatial priors, and a Boundary Refinement Module (BRM) designed to improve edge delineation. The authors evaluate their method on the ACDC benchmark, reporting a Dice score of 93.39%, which they claim outperforms existing benchmarks and exceeds human expert consistency.

Significance. If the performance gains are robustly validated, this work represents a successful application of 'foundation model' distillation for medical tasks where data is scarce. The methodology of freezing a large encoder and focusing on domain-specific decoder modules is an efficient paradigm. However, the significance is currently tempered by questions regarding the baseline comparisons and the robustness of the spatial priors used in the decoder.

major comments (3)

[Table 1, Performance Comparison] The reported performance for the nnU-Net baseline (89.50% Dice) is substantially lower than established benchmarks for the ACDC dataset. Standard implementations of nnU-Net typically achieve Dice scores between 91.5% and 92.5% on this specific dataset (e.g., Isensee et al., 2021). The claim of a +3.89% improvement is likely inflated by an under-performing baseline. The authors must either provide a justification for this low baseline score or re-evaluate against a properly tuned state-of-the-art configuration to substantiate the claim of superiority.
[§3.2, Eq. (1) and (2)] The 'Cardiac-Specific Attention' module is described as 'topology-aware,' yet it utilizes a template T defined as the arithmetic mean of training masks. This is a spatial location prior, not a topological one. Because this template is fixed in image space, it assumes the heart is consistently centered and scaled, as it is in the ACDC dataset. The authors should evaluate the model's sensitivity to translation and rotation (e.g., by jittering the input images) to determine if the module generalizes beyond rigidly aligned datasets or if it simply overfits the ACDC centering convention.
[§3.3, Boundary Refinement Module] The BRM relies on 'edge features' to refine the segmentation. However, the manuscript does not specify how the ground truth for these edges is defined during training or how the boundary loss is weighted relative to the primary Dice/CE loss. Without this information, the reproducibility of the refinement process is compromised.

minor comments (3)

[§3.1, Architecture Overview] The authors state the SAM encoder is 'frozen.' It would be beneficial to clarify if any adapters or prompt-based tuning were attempted, as recent literature (e.g., MedSAM) suggests freezing the encoder entirely can limit the capture of medical-specific textures.
[Figure 2, Qualitative Results] The visual comparison would be more informative if it included error maps (residuals) between the ground truth and the predictions, particularly at the myocardial boundaries where the BRM is claimed to provide the most benefit.
[Abstract / Introduction] The term 'topological prior' is used throughout but is mathematically imprecise here. 'Spatial prior' or 'anatomical atlas prior' would be more accurate given the implementation in Section 3.2.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive critique of CardioSAM. We particularly appreciate the recognition of the efficiency of our 'frozen encoder' paradigm. We acknowledge the valid points regarding the baseline comparison on the ACDC dataset and the precision of our 'topology-aware' terminology. In our revision, we will update the performance benchmarks to include state-of-the-art ensemble nnU-Net results, provide a sensitivity analysis regarding the spatial priors used in the CSA module, and include the technical implementation details for the Boundary Refinement Module (BRM). We believe these changes significantly strengthen the manuscript's empirical rigor and clarity.

read point-by-point responses

Referee: [Table 1, Performance Comparison] The reported performance for the nnU-Net baseline (89.50% Dice) is substantially lower than established benchmarks for the ACDC dataset. Standard implementations of nnU-Net typically achieve Dice scores between 91.5% and 92.5% on this specific dataset... The authors must either provide a justification for this low baseline score or re-evaluate against a properly tuned state-of-the-art configuration to substantiate the claim of superiority.

Authors: The referee is correct that ensemble-based or highly-tuned versions of nnU-Net reach significantly higher scores (91.5%–92.5%) on ACDC than our reported 89.50%. The lower figure in our manuscript was obtained from a vanilla, single-fold implementation. We acknowledge that this comparison is not representative of the true state-of-the-art performance. In the revised manuscript, we will update Table 1 to include the 92.5% benchmark from Isensee et al. (2021). While the improvement margin over nnU-Net will be reduced from +3.89% to approximately +0.9-1.3%, our model's result of 93.39% still demonstrates a clear improvement, even against a properly tuned baseline. revision: yes
Referee: [§3.2, Eq. (1) and (2)] The 'Cardiac-Specific Attention' module is described as 'topology-aware,' yet it utilizes a template T defined as the arithmetic mean of training masks. This is a spatial location prior, not a topological one... The authors should evaluate the model's sensitivity to translation and rotation (e.g., by jittering the input images) to determine if the module generalizes beyond rigidly aligned datasets or if it simply overfits the ACDC centering convention.

Authors: We appreciate this terminological clarification. While 'topology' was intended to refer to the invariant relational arrangement of the cardiac chambers, we agree that 'anatomical-spatial prior' is a more accurate description of the template-based approach. To address the concern regarding sensitivity to centering, we have conducted robustness experiments involving image jittering (random translation of +/- 10% and rotation of +/- 15 degrees). Results indicate that the model maintains a Dice score above 92.8%, likely because the rich features from the frozen SAM encoder provide sufficient contextual information to handle minor misalignments with the spatial template. We will update Section 3.2 with this revised terminology and the robustness analysis. revision: partial
Referee: [§3.3, Boundary Refinement Module] The BRM relies on 'edge features' to refine the segmentation. However, the manuscript does not specify how the ground truth for these edges is defined during training or how the boundary loss is weighted relative to the primary Dice/CE loss. Without this information, the reproducibility of the refinement process is compromised.

Authors: We apologize for the omission of these technical specifications. The edge ground truth is generated on-the-fly by applying a morphological gradient (dilation minus erosion) to the binary ground-truth segmentation masks, creating a boundary mask approximately 2 pixels wide. The total training loss is defined as L = L_Dice + L_CE + 0.5 * L_edge, where L_edge is a binary cross-entropy loss applied to the predicted edge map. These details, along with the specific network layers used to extract the edge features, will be included in the revised Section 3.3. revision: yes

Circularity Check

2 steps flagged

CardioSAM's 'Topological' innovation is a spatial average mask (atlas) that effectively smuggles training label locations into the prediction phase.

specific steps

fitted input called prediction [Section 3.2, Equation 1 & 2]
"The CSA module incorporates anatomical topological priors... We define an anatomical template T where T is calculated by averaging the ground-truth segmentation masks across the training set. F_out = Softmax((Q*K^T)/sqrt(d_k) * T)."

The paper claims to derive 'topology-aware' predictions. However, the 'topological prior' T is explicitly the mean of the training labels (an atlas). By multiplying the attention weights by T, the model is mathematically restricted to predicting the heart in the average location of the training set. On a centered dataset like ACDC, the high Dice score is partially a result of the model being 'given' the spatial answer via this fitted template, rather than learning topology from first principles.
renaming known result [Section 3.2]
"The proposed decoder introduces two key innovations: a Cardiac-Specific Attention module that incorporates anatomical topological priors..."

The paper renames a standard spatial probability map (or 'atlas') as a 'topological prior.' Topology refers to connectivity and relative properties invariant to distance; an average spatial mask is a coordinate-dependent rigid prior. By framing a spatial coordinate constraint as a 'topological innovation,' the paper presents a common pre-processing/atlas-weighting step as a novel structural derivation.

full rationale

The circularity in CardioSAM arises from the 'Cardiac-Specific Attention' (CSA) module. The paper defines 'topological priors' not as geometric invariants (like Euler characteristics or Betti numbers), but as a voxel-wise average of the training masks ($T$). This template $T$ is then used as a hard gate (multiplier) on the attention mechanism. Because the ACDC dataset is highly standardized and centered, the 'topological prior' is essentially a spatial cheat-sheet that tells the model exactly where the classes are expected to be based on the labels it was trained on. While the use of a frozen SAM encoder provides some independent feature extraction, the +3.89% gain over the baseline is heavily influenced by this spatial fit and a baseline (nnU-Net at 89.5%) that is significantly lower than established benchmarks (92%+), creating a narrative of 'innovation' through a combination of a rigid prior and a weak comparator.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The model relies on the standard deep learning axiom that feature representations from general datasets are transferable to specific medical domains via specialized decoders.

free parameters (1)

Cardiac-Specific Attention weights
These are the primary learned parameters in the custom decoder.

axioms (1)

domain assumption SAM frozen encoder features are sufficient for CMR
The model relies on the pre-trained weights of SAM being relevant to the visual textures of MRI scans.

invented entities (1)

Cardiac-Specific Attention module no independent evidence
purpose: To incorporate anatomical topological priors into the decoding process.
This is a new architectural component introduced by the authors.

pith-pipeline@v0.9.0 · 6285 in / 1530 out tokens · 14201 ms · 2026-05-08T02:24:09.483436+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

25 extracted references · 4 canonical work pages · 2 internal anchors

[1]

Deep learning for cardiovascular imaging: From research to clinical practice,

D. C. Peters, L. Axel, G. Captur, H. El-Rewaidy, M. Gatti, O. Gjesdal, D. N. Metaxas, D. Ouyang, S. E. Petersen, M. Popet al., “Deep learning for cardiovascular imaging: From research to clinical practice,”JACC: Cardiovascular Imaging, vol. 16, no. 2, pp. 261–279, 2023

2023
[2]

Deep learning for cardiac image segmentation: A review,

C. Chen, C. Qin, H. Qiu, G. Tarroni, J. Duan, W. Bai, and D. Rueckert, “Deep learning for cardiac image segmentation: A review,”Frontiers in Cardiovascular Medicine, vol. 7, p. 25, 2020

2020
[3]

Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: a benchmark,

O. Bernard, A. Lalande, C. Zotti, F. Cervenansky, X. Yang, P.-A. Heng, I. Cetin, K. Lekadir, O. Camara, M. A. Gonzalez Ballesteret al., “Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: a benchmark,”IEEE transactions on medical imaging, vol. 37, no. 9, pp. 2059–2073, 2018

2059
[4]

Deep learning approaches to biomedical image segmentation,

I. R. I. Haque and J. Neubert, “Deep learning approaches to biomedical image segmentation,”Informatics in Medicine Unlocked, vol. 18, p. 100297, 2020

2020
[5]

Deep learning for medical image segmentation: State-of-the-art advancements and challenges,

M. E. Rayed, S. S. Islam, S. I. Niha, J. R. Jim, M. M. Kabir, and M. F. Mridha, “Deep learning for medical image segmentation: State-of-the-art advancements and challenges,”Informatics in Medicine Unlocked, vol. 47, p. 101504, 2024

2024
[6]

U-net: Convolutional networks for biomedical image segmentation,

O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” inMedical Image Computing and Computer-Assisted Intervention – MICCAI 2015. Springer, 2015, pp. 234–241

2015
[7]

TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation

J. Chen, Y . Lu, Q. Yu, X. Luo, E. Adeli, Y . Wang, L. Lu, A. L. Yuille, and Y . Zhou, “Transunet: Transformers make strong encoders for medical image segmentation,”arXiv preprint arXiv:2102.04306, 2021

work page internal anchor Pith review arXiv 2021
[8]

Segment anything,

A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo, P. Doll ´ar, and R. Girshick, “Segment anything,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2023, pp. 3879–3890

2023
[9]

Medsam: Segment anything in medical images,

J. He, C. Chen, Y . Li, J. Wang, A. P. Yu, X. Li, A. L. Yuille, and Y . Zhou, “Medsam: Segment anything in medical images,”arXiv preprint arXiv:2304.09324, 2023

work page arXiv 2023
[10]

Segment anything in medical images,

J. Ma, Y . He, F. Li, L. Yang, B.-k. Zhu, J.-w. Bai, R.-x. Wang, X.-p. Zhang, R.-q. Liu, C. Liet al., “Segment anything in medical images,”arXiv preprint arXiv:2304.14660, 2023

work page arXiv 2023
[11]

Simple python module for conversions between dicom images and radiation therapy structures, masks, and prediction arrays,

B. M. Anderson, K. A. Wahid, and K. K. Brock, “Simple python module for conversions between dicom images and radiation therapy structures, masks, and prediction arrays,”Practical radiation oncology, vol. 11, no. 3, pp. 226–229, 2021

2021
[12]

Encoder-decoder with atrous separable convolution for semantic image segmentation,

L.-C. Chen, Y . Zhu, G. Papandreou, F. Schroff, and H. Adam, “Encoder-decoder with atrous separable convolution for semantic image segmentation,” inProceedings of the European conference on computer vision (ECCV), 2018, pp. 801–818

2018
[13]

nnu-net: Self-adapting framework for u-net-based medical image segmentation,

F. Isensee, J. Petersen, A. Klein, D. Zimmerer, P. F. Jaeger, S. Kohl, J. Wasserthal, G. Koehler, T. Norajitra, S. Wirkert, and K. H. Maier-Hein, “nnu-net: Self-adapting framework for u-net-based medical image segmentation,” inBildverarbeitung f ¨ur die Medizin 2019. Springer Fachmedien Wiesbaden, 2019

2019
[14]

Statistical validation of image segmentation quality based on a spatial overlap index1: scientific reports,

K. H. Zou, S. K. Warfield, A. Bharatha, C. M. C. Tempany, M. R. Kaus, S. J. Haker, W. M. Wells, F. A. Jolesz, and R. Kikinis, “Statistical validation of image segmentation quality based on a spatial overlap index1: scientific reports,”Academic Radiology, vol. 11, no. 2, pp. 178–189, 2004

2004
[15]

Generalized intersection over union: A metric and a loss for bounding box regression,

H. Rezatofighi, N. Tsoi, J. Gwak, A. Sadeghian, I. Reid, and S. Savarese, “Generalized intersection over union: A metric and a loss for bounding box regression,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2019, pp. 658–666

2019
[16]

Heart failure and ejection fraction,

S. Hajouli and D. Ludhwani, “Heart failure and ejection fraction,”StatPearls [Internet], 2020

2020
[17]

Evaluation of right ventricular volume and systolic function in normal fetuses using intelligent spatiotemporal image correlation,

J.-X. Sun, A.-L. Cai, and L.-M. Xie, “Evaluation of right ventricular volume and systolic function in normal fetuses using intelligent spatiotemporal image correlation,”World Journal of Clinical Cases, vol. 7, no. 15, pp. 2003–2011, 2019

2003
[18]

Explainable ai in cardiac imaging: A state-of-the-art review,

M. Salih, A. Al-antali, M. Al-Jefri, and M. H. Al-Mallah, “Explainable ai in cardiac imaging: A state-of-the-art review,”Circulation: Cardiovascular Imaging, vol. 16, no. 4, p. e014519, 2023

2023
[19]

From blackbox to trustworthy: A review on explainable and trustworthy ai in medical image segmentation,

L. Wang, Y . Li, S. Wang, X. Li, S. Wang, X. Li, Y . Li, X. Li, Y . Li, X. Liet al., “From blackbox to trustworthy: A review on explainable and trustworthy ai in medical image segmentation,”Journal of Healthcare Engineering, vol. 2024, 2024

2024
[20]

Deep learning based multi-modal cardiac mr image segmentation,

N. Karani, K. Chaitanya, C. Baumgartner, and E. Konukoglu, “Deep learning based multi-modal cardiac mr image segmentation,” inInternational Workshop on Statistical Atlases and Computational Models of the Heart. Springer, 2019, pp. 139–148

2019
[21]

Disentangling modality and anatomy for cardiac image segmentation,

A. Ghavami, N. Karani, and E. Konukoglu, “Disentangling modality and anatomy for cardiac image segmentation,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2019, pp. 443–451

2019
[22]

Distilling the Knowledge in a Neural Network

G. Hinton, O. Vinyals, and J. Dean, “Distilling the knowledge in a neural network,”arXiv preprint arXiv:1503.02531, 2015

work page internal anchor Pith review arXiv 2015
[23]

Knowledge distillation: A survey,

J. Gou, B. Yu, S. J. Maybank, and D. Tao, “Knowledge distillation: A survey,”International Journal of Computer Vision, vol. 129, no. 6, pp. 1789–1819, 2021

2021
[24]

A blinded, randomized trial of an artificial intelligence system for the assessment of cardiac function,

D. Ouyang, B. He, A. Ghorbani, N. Yuan, J. Ebinger, C. P. Langlotz, P. A. Heidenreich, R. A. Harrington, D. H. Liang, E. A. Ashley, and J. Y . Zou, “A blinded, randomized trial of an artificial intelligence system for the assessment of cardiac function,”Nature, vol. 616, no. 7956, pp. 333–338, 2023

2023
[25]

Artificial intelligence to accelerate evidence generation from clinical trials,

N. R. Desai and J. S. Ross, “Artificial intelligence to accelerate evidence generation from clinical trials,”Journal of the American College of Cardiology, vol. 83, no. 10, pp. 994–1006, 2024

2024