REVIEW 4 major objections 5 minor 16 references
Multi-Scale Feature Fusion with Image-Driven Spatial Integration for Left Atrium Segmentation from Cardiac MRI Images
T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash
Pith's one-line read A frozen DINOv2 encoder, paired with a UNet-style decoder, reaches 92.3% Dice and 84.1% IoU on LAScarQS 2022 left atrium segmentation, beating the nnUNet baseline by 4.6 and 2.4 percentage points.
desk verdict Plausible incremental result on left atrium segmentation, but the advertised learnable weighting is never actually ablated—the headline numbers don't match the ablation table. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the learnable multi-scale feature fusion: each encoder block feature $F_i$ is assigned a weight $w_i=\exp(\theta_i)/\sum_j\exp(\theta_j)$ and the blocks are aggregated as $\hat F=\sum_i w_i F_i$, with the top-$k$ weighted blocks fed to the decoder. Alongside it, each decoder stage concatenates the transformed original image $I'$, the encoder feature $F_i$, and the previous decoder output $D_{j-1}$ to form $F_{\mathrm{concat}}=\mathrm{concat}(I', F_i, D_{j-1})$. Together these mechanisms let a frozen patch-based transformer preserve the pixel-level boundary detail it otherwise discards during downsampling.
What would settle it
Re-run the comparison with patient-level 5-fold cross-validation on LAScarQS 2022 and report the mean and confidence interval of the Dice gap; if the giant DINOv2 model does not beat nnUNet in a majority of folds, or the lower confidence bound of the gap includes zero, the central claim fails.
Extended reading notes
Core claim
The paper claims that DINOv2, a vision transformer pretrained on natural images, can be repurposed as a frozen encoder for medical image segmentation if the decoder is designed to compensate for its low-resolution patch-based features. The proposed framework aggregates encoder block outputs with normalized learnable weights, feeds the original input image back into each decoder stage, and uses a UNet-style decoder to recover spatial detail. On the LAScarQS 2022 left atrium task, the giant DINOv2 configuration reaches 92.3% mean Dice and 84.1% mean IoU, beating the nnUNet baseline by 4.6 and 2.4 percentage points respectively, with the giant model's improvement reported as statistically significant at p<0.05.
Load-bearing premise
The load-bearing premise is that a patient-level 20% test split of the 60 labeled volumes is large enough for the reported Dice and IoU gaps over nnUNet to be statistically reliable rather than split-dependent noise.
Editorial extensions
If this is right
- Frozen DINOv2 encoders with a suitable decoder can exceed task-specific baselines like nnUNet on left atrium MRI, suggesting that pretraining on natural images transfers to small medical datasets.
- Learnable block weighting provides a principled alternative to hand-selecting encoder depths for dense prediction tasks, and the ablation suggests the choice of blocks materially changes Dice.
- Re-introducing the raw input image at each decoder stage restores boundary detail lost to patch-based downsampling, a design that can be ported to other segmentation architectures.
- The reported 92.3% Dice performance brings automated left atrium segmentation closer to clinically useful levels for atrial fibrillation treatment planning, provided the result survives external validation.
Reading between the lines
- A natural follow-up is to inspect the learned block weights after training: if the largest weights concentrate on mid-depth blocks, that would explain why the hand-picked [7,5,3,1] selection worked well and could guide encoder choices for other organs.
- The same decoder recipe could be applied to other frozen vision transformers or to 3D cardiac MRI, where the encoder's resolution loss is even more severe and the fusion mechanism would be tested more aggressively.
- Because the test set is only about 12 volumes, the p<0.05 result is preliminary; an external dataset with more patients would be the strongest check on whether the Dice gain reproduces.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript proposes a left-atrium segmentation framework that combines a frozen DINOv2 vision-transformer encoder with a UNet-style decoder, a learnable softmax block-weighting scheme for multi-scale feature fusion, and input-image integration at each decoder stage. The method is evaluated on the LAScarQS 2022 left-atrium dataset of 60 labeled LGE-MRI volumes, with a 70/10/20 patient-level split. Table I reports mean Dice values of 91.5%, 92.1%, and 92.3% for DINOv2-base, large, and giant respectively, against 87.7% for a 2D nnUNet baseline, with only the giant model marked as statistically significant (p<0.05). The ablation study in Table II reports Dice values between 87.1% and 90.3% for two fixed block-selection choices, with and without spatial integration. The authors attribute the reported gains to the learnable multi-scale feature fusion and input-image integration.
Significance. If the proposed configuration indeed yields a 92.3% Dice score on LAScarQS 2022, the result would be practically useful for automated left-atrium segmentation. The paper is concise, uses a public benchmark, reports standard deviations, and provides explicit equations for the weighting and fusion operations. The main contribution, however, is not currently supported by the experiments: the learnable weighting mechanism is described formally but never isolated in an ablation, and the headline result is not reproduced by any configuration shown in the ablation table. Because the central claim rests on an unverified attribution of the improvement to a specific component, the significance of the paper in its present form is limited.
major comments (4)
- [Section IV, Table II vs. Section III, Table I] The ablation table does not contain a configuration that reproduces the headline results of Table I. The best Dice score in Table II is 90.3% (DINOv2-giant, blocks [7,5,3,1], with spatial integration), which is 2.0 points below the 92.3% reported for DINOv2-giant in Table I. This mismatch makes it impossible to determine which architecture produced the main result. Please specify the exact configuration used for Table I and include that configuration in the ablation study, or state explicitly if Table I was obtained with a different setup.
- [Section II-B, Eqs. (1)-(2); Section IV, Table II] The proposed learnable block weighting is never ablated. Table II tests only fixed block selections ('Last 4 blocks' and '[7,5,3,1]') with and without spatial integration, so the contribution of the learnable softmax weighting to the reported accuracy is unknown. Because the abstract and conclusions credit this mechanism for the improvement, please add experiments that compare the learned weighting (Eqs. (1)-(2)) against the fixed block selections, and report the learned weights or the selected k for each model.
- [Section II-B, Eqs. (1)-(2)] The method description is internally inconsistent: Eq. (1) aggregates feature maps from all N encoder blocks into a single tensor, while the following paragraph states that only the top k blocks are selected for the decoder and that 'this approach retains all encoder blocks.' Please clarify whether non-selected blocks are zeroed, whether Eq. (1) is applied before or after block selection, how k is chosen, and how the learnable weights interact with the top-k selection.
- [Section III, Table I; Section V] The statistical evidence is weaker than the text suggests. Only DINOv2-giant is marked as significant (p<0.05); DINOv2-base and DINOv2-large are not marked significant, yet the conclusions state that the method achieved 'significant improvements in Dice and IoU scores' without qualification. In addition, with roughly 12 test volumes and standard deviations of about 5-6 Dice points, the significance of the giant model should be reported with the exact test type, confidence intervals, and effect sizes. Please also report whether the comparison to nnUNet is paired or unpaired.
minor comments (5)
- [Section II-F] The paper does not state how the 3D LGE-MRI volumes are converted into the 2D inputs used by DINOv2, which slices were used, how normalization was performed, or whether any data augmentation was applied. These details are needed for reproducibility.
- [Section II-F] Reference [15] is Loshchilov and Hutter's 'Decoupled weight decay regularization' (AdamW), but the text says the Adam optimizer was used. Please clarify whether the optimizer is Adam or AdamW and cite accordingly.
- [Fig. 1 caption] The caption says 'input image augmentation,' but the described mechanism is input-image integration into the decoder. Please correct the terminology for consistency with Section II-C.
- [Table II] The column headers 'Selected Blocksfrom Encoder' and 'Spatial Integration' are merged and hard to read. Please reformat the table for clarity.
- [Section II-A] In the phrase 'D, H and W are the dimension, height, and weight of the embedding of the image,' the word 'weight' should be 'width.'
Circularity Check
No circularity: the reported scores are measured on a held-out patient split, and no load-bearing claim reduces to fitted values or to a self-citation.
full rationale
The paper's central claim is an empirical performance comparison: the proposed DINOv2-based framework is trained and evaluated on a patient-level split of the LAScarQS 2022 dataset, with 20% reserved for testing. The reported Dice and IoU values are computed on this held-out test set and are not derived algebraically from any fitted parameter or from the learnable block weights. Equation (1) and (2) define the weighted feature aggregation, but these weights are trained on the training portion and are not used in the evaluation as a restatement of the final metric; therefore there is no fitted-input-called-prediction step. The only self-citation, reference [10], is used in Section II merely to describe the authors' previous investigation with linear decoders and is not load-bearing for the claimed improvement. The discrepancies between the headline result in Table I and the ablation configurations in Table II, as well as the unspecified value of k in the top-k selection, are internal-consistency and reproducibility concerns rather than circularity. No equation or citation chain reduces the paper's conclusions to its own inputs, so the circularity burden is not met.
Assumptions & free parameters
free parameters (3)
- Encoder block selection [7,5,3,1] =
indices 7, 5, 3, 1
- Input image size =
448x448
- Top-k (number of selected encoder blocks) =
not reported
assumptions (3)
- domain assumption The LAScarQS 2022 dataset provides a representative sample of LGE-MRI left atrium anatomy across scanners and patients.
- domain assumption DINOv2 features pretrained on natural images transfer to LGE-MRI with a frozen encoder.
- domain assumption Slicing 3D volumes into independent 2D images preserves the anatomical context needed for accurate LA segmentation.
Cite this review
Pith. "Pith review of Multi-Scale Feature Fusion with Image-Driven Spatial Integration for Left Atrium Segmentation from Cardiac MRI Images." pith.science (2026). https://pith.science/paper/USB42DMX
@misc{pith2026250206615,
author = {Pith},
title = {Pith review of: Multi-Scale Feature Fusion with Image-Driven Spatial Integration for Left Atrium Segmentation from Cardiac MRI Images},
year = {2026},
howpublished = {\url{https://pith.science/paper/USB42DMX}},
note = {Machine review of arXiv:2502.06615}
}
read the original abstract
Accurate segmentation of the left atrium (LA) from late gadolinium-enhanced magnetic resonance imaging plays a vital role in visualizing diseased atrial structures, enabling the diagnosis and management of cardiovascular diseases. It is particularly essential for planning treatment with ablation therapy, a key intervention for atrial fibrillation (AF). However, manual segmentation is time-intensive and prone to inter-observer variability, underscoring the need for automated solutions. Class-agnostic foundation models like DINOv2 have demonstrated remarkable feature extraction capabilities in vision tasks. However, their lack of domain specificity and task-specific adaptation can reduce spatial resolution during feature extraction, impacting the capture of fine anatomical detail in medical imaging. To address this limitation, we propose a segmentation framework that integrates DINOv2 as an encoder with a UNet-style decoder, incorporating multi-scale feature fusion and input image integration to enhance segmentation accuracy. The learnable weighting mechanism dynamically prioritizes hierarchical features from different encoder blocks of the foundation model, optimizing feature selection for task relevance. Additionally, the input image is reintroduced during the decoding stage to preserve high-resolution spatial details, addressing limitations of downsampling in the encoder. We validate our approach on the LAScarQS 2022 dataset and demonstrate improved performance with a 92.3% Dice and 84.1% IoU score for giant architecture compared to the nnUNet baseline model. These findings emphasize the efficacy of our approach in advancing the field of automated left atrium segmentation from cardiac MRI.
Figures
Reference graph
Works this paper leans on
-
[1]
Automated left atrial time-resolved segmentation in mri long-axis cine images using active contours,
R. A. Gonzales, F. Seemann, J. Lamy, P. M. Arvidsson, E. Heiberg, V . Murray, and D. C. Peters, “Automated left atrial time-resolved segmentation in mri long-axis cine images using active contours,” BMC Medical Imaging , vol. 21, no. 1, p. 101, 2021
work page 2021
-
[2]
Left atrial segmentation challenge: a unified benchmarking framework,
C. Tobon-Gomez, J. Peters, J. Weese, K. Pinto, R. Karim, T. Schaeffter, R. Razavi, and K. S. Rhode, “Left atrial segmentation challenge: a unified benchmarking framework,” in Statistical Atlases and Compu- tational Models of the Heart. Imaging and Modelling Challenges: 4th International Workshop, STACOM 2013, Held in Conjunction with MICCAI 2013, Nagoya, ...
work page 2013
-
[3]
Lassnet: A four steps deep neural network for left atrial segmentation and scar quantification,
A. L. Lefebvre, C. A. Yamamoto, J. K. Shade, R. P. Bradley, R. A. Yu, R. L. Ali, D. M. Popescu, A. Prakosa, E. G. Kholmovski, and N. A. Trayanova, “Lassnet: A four steps deep neural network for left atrial segmentation and scar quantification,” in Challenge on Left Atrial and Scar Quantification and Segmentation , pp. 1–15, Springer, 2022
work page 2022
-
[4]
Deep u-net architecture with curriculum learning for left atrial segmentation,
L. Jiang, Y . Li, Y . Wang, H. Cui, Y . Xia, and Y . Zhang, “Deep u-net architecture with curriculum learning for left atrial segmentation,” in Challenge on Left Atrial and Scar Quantification and Segmentation , pp. 115–123, Springer, 2022
work page 2022
-
[5]
M. Wu, D. Zhang, Y . Hua, M. Si, P. Liu, and Q. Wang, “Transfusion: Efficient vision transformer based on 3d transesophageal echocardio- graphy images for the left atrial appendage segmentation,” Expert Systems with Applications , vol. 255, p. 124727, 2024
work page 2024
-
[6]
Us- former: A small network for left atrium segmentation of 3d lge mri,
H. Lin, S. L ´opez-Tapia, F. Schiffers, Y . Wu, S. Gunasekaran, J. Hwang, D. Bishara, E. Kholmovski, M. Elbaz, R. S. Passman, et al. , “Us- former: A small network for left atrium segmentation of 3d lge mri,” Heliyon, vol. 10, no. 7, 2024
work page 2024
-
[7]
Mlc: Multi-level consistency learning for semi-supervised left atrium segmentation,
Z. Shi, M. Jiang, Y . Li, B. Wei, Z. Wang, Y . Wu, T. Tan, and G. Yang, “Mlc: Multi-level consistency learning for semi-supervised left atrium segmentation,” Expert Systems with Applications , vol. 244, p. 122903, 2024
work page 2024
-
[8]
A generalizable 3D framework and model for self-supervised learning in medical imaging
T. Xu, S. Hosseini, C. Anderson, A. Rinaldi, R. G. Krishnan, A. L. Martel, and M. Goubran, “A generalizable 3d framework and model for self-supervised learning in medical imaging,” arXiv preprint arXiv:2501.11755, 2025
work page Pith review arXiv 2025
Show all 16 references
-
[9]
Towards general purpose vision foundation models for medical image analysis: An experimental study of dinov2 on radiology benchmarks,
M. Baharoon, W. Qureshi, J. Ouyang, Y . Xu, K. Phol, A. Aljouie, and W. Peng, “Towards general purpose vision foundation models for medical image analysis: An experimental study of dinov2 on radiology benchmarks,” arXiv preprint arXiv:2312.02366 , 2023
2023 arXiv
-
[10]
Assessing the performance of the dinov2 self-supervised learning vision transformer model for the segmentation of the left atrium from mri images,
B. Kundu, B. Khanal, R. Simon, and C. A. Linte, “Assessing the performance of the dinov2 self-supervised learning vision transformer model for the segmentation of the left atrium from mri images,” arXiv preprint arXiv:2411.09598, 2024
2024 arXiv
-
[11]
Dinov2: Learning robust visual features without supervision,
M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khali- dov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby, et al., “Dinov2: Learning robust visual features without supervision,” arXiv preprint arXiv:2304.07193, 2023
2023 arXiv
-
[12]
U-net: Convolutional networks for biomedical image segmentation,
O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, pro- ceedings, part III 18 ...
2015
-
[13]
Left atrial and scar quantification & segmentation challenge
LaScarQS, “Left atrial and scar quantification & segmentation challenge.” 2022, https://zmiclab.github.io/projects/ lascarqs22. (Accessed: 20 January 2025)
2022
-
[14]
Research computing services,
R. I. of Technology, “Research computing services,” 2019
2019
-
[15]
Decoupled weight decay regularization,
I. Loshchilov, “Decoupled weight decay regularization,” arXiv preprint arXiv:1711.05101, 2017
2017 arXiv
-
[16]
nnu-net: a self-configuring method for deep learning-based biomedical image segmentation,
F. Isensee, P. F. Jaeger, S. A. Kohl, J. Petersen, and K. H. Maier- Hein, “nnu-net: a self-configuring method for deep learning-based biomedical image segmentation,” Nature methods , vol. 18, no. 2, pp. 203–211, 2021
2021
Reviewed August 8, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.