arxiv: 2605.03602 · v1 · submitted 2026-05-05 · 📡 eess.IV

Recognition: unknown

Dante: An Open Source Model Pre-Training and Fine-Tuning Tool for the Dafne Federated Framework for Medical Image Segmentation

Dibya Kumari, Francesco Santini, Giuseppe Timpano, Maria Eugenia Caligiuri

Pith reviewed 2026-05-07 12:54 UTC · model grok-4.3

classification 📡 eess.IV

keywords DanteDafneLoRAmedical image segmentationfine-tuningMRIgradual unfreezingfederated learning

0 comments

The pith

Dante provides an open-source fine-tuning backend that adapts pre-trained medical segmentation models to new MRI domains using gradual unfreezing and LoRA.

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

Adapting pre-trained deep learning models for medical image segmentation to new clinical domains remains difficult when annotated data at the target site are scarce. Dante integrates with the Dafne federated ecosystem as a dedicated training module that supports from-scratch training with automatic architecture configuration, configurable layer freezing schedules, and an extension of Low-Rank Adaptation to N-dimensional convolutional layers through channel-wise factorization. Validation on realistic cross-domain MRI transfers for abdominal organ segmentation and brain white matter lesion segmentation, under both full-data and few-shot conditions, shows gradual unfreezing reaching 85 percent of peak performance in up to 63.6 percent fewer epochs and LoRA attaining Dice scores up to 0.957. Both strategies outperform the baseline, with gains increasing when the source pre-training dataset is richer. These results position Dante as a domain-agnostic fine-tuning module suited to real clinical deployment.

Core claim

Dante serves as the training and fine-tuning backend for the Dafne federated segmentation framework, enabling training from scratch via automatic architecture setup along with two adaptation strategies: gradual unfreezing of layers and Low-Rank Adaptation extended to N-dimensional convolutions by channel-wise factorization. Across cross-domain MRI scenarios covering abdominal organs and brain white matter lesions in full-data and few-shot regimes, these approaches consistently yield faster convergence and higher final accuracy than standard training from scratch, with the magnitude of improvement scaling according to the richness of the initial pre-training data.

What carries the argument

Dante's LoRA extension to N-dimensional convolutional layers via channel-wise factorization together with gradual unfreezing schedules, which allow selective parameter updates while retaining prior representations during domain transfer.

If this is right

Gradual unfreezing shortens the number of epochs needed to reach high segmentation performance when moving to a new domain.
LoRA delivers strong Dice coefficients while updating only a small subset of parameters, even in few-shot settings.
Larger performance gains occur when the source pre-training dataset is more extensive or diverse.
The same adaptation methods succeed across different anatomical targets such as organs and lesions.
Dante functions as a reusable, domain-agnostic component inside the Dafne federated framework.

Where Pith is reading between the lines

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

New clinical sites could adapt existing models with substantially less local annotation effort.
The federated integration may allow collaborative adaptation across institutions while keeping raw patient data private.
The channel-wise LoRA factorization could be applied to other volumetric modalities such as CT or PET.
Faster domain adaptation may accelerate practical deployment of segmentation tools in varied hospital environments.

Load-bearing premise

The specific cross-domain MRI transfer scenarios for abdominal organs and brain lesions, plus the full-data versus few-shot conditions tested, are representative of real clinical deployment and that the observed performance gains generalize beyond the datasets and architectures used.

What would settle it

A new experiment at an independent clinical site with a different MRI scanner and patient population in which neither gradual unfreezing nor LoRA matches or exceeds the Dice scores and convergence speed of standard fine-tuning on the same segmentation tasks.

Figures

Figures reproduced from arXiv: 2605.03602 by Dibya Kumari, Francesco Santini, Giuseppe Timpano, Maria Eugenia Caligiuri.

**Figure 3.** Figure 3: Dafne Model Trainer GUI - an example of training from scratch using the CHAOS dataset. During both training from scratch and fine-tuning, the GUI displays validation loss curves and additional information, such as the best average dice score and the current epoch's dice score for each segmented organ view at source ↗

read the original abstract

Adapting pre-trained deep learning segmentation models to new clinical domains is a persistent challenge in medical image analysis, particularly when annotated data at the target site are scarce. Parameter-efficient fine-tuning strategies offer a principled solution by selectively updating a controlled subset of model parameters, preserving previously acquired representations while reducing the risk of overfitting on small datasets. This paper introduces DAfNe TrainEr (Dante), an open-source module integrating with the Dafne federated segmentation ecosystem as a dedicated training and fine-tuning backend. Dante supports training from scratch with automatic architecture configuration, configurable layer freezing schedules, and Low-Rank Adaptation (LoRA) extended to N-dimensional convolutional layers through channel-wise factorization. To validate the module, Gradual Unfreezing (GU) and LoRA are assessed across realistic cross-domain MRI transfer scenarios covering abdominal organ segmentation and brain white matter lesion segmentation, under full-data and few-shot conditions. GU reduced the epochs required to reach 85% of peak performance by up to 63.6% compared to training from scratch, while LoRA achieved Dice Similarity Coefficients up to 0.957 in data-rich scenarios. Both strategies outperformed the baseline across all tested domains, with gains amplified by richer pre-training datasets. These results validate Dante as a domain-agnostic fine-tuning module for medical image segmentation in real clinical deployment conditions. Dante code is available at https://github.com/dafne-imaging/dafne-torch-trainer while Dafne ecosystem project is available at https://github.com/dafne-imaging.

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

Dante is a practical open-source trainer for the Dafne framework that shows efficiency gains on two MRI tasks, but its domain-agnostic clinical claims rest on narrow evidence.

read the letter

The main thing to know about this paper is that it delivers a practical open-source fine-tuning tool called Dante for the Dafne framework, with experiments showing efficiency gains on two MRI segmentation tasks. What is new is the integration of gradual unfreezing and a custom N-dimensional LoRA using channel-wise factorization, all packaged as a backend for pre-training and adaptation in a federated medical imaging system. The code is released on GitHub, which is a real plus for reproducibility. The paper does well in laying out how these parameter-efficient methods apply to the existing Dafne setup and in providing quantitative results across full-data and few-shot conditions for abdominal organ and brain lesion segmentation. The reported reductions in training epochs and the high Dice scores give a sense that the approach can help when target data is limited. The soft spots come from the narrow experimental base. All validation uses only MRI data on those two specific tasks, with no tests involving other modalities, protocol variations, or actual federated multi-site training that would exercise the privacy features of Dafne. This makes the assertion that the results validate domain-agnostic fine-tuning for real clinical deployment conditions feel premature. The abstract also does not include enough on the precise baselines, dataset details, or significance testing, though the full paper might fill that in. These are not fatal but they limit how far the conclusions can be pushed. This paper is aimed at researchers and developers building or using federated tools for medical image segmentation who want a ready-made trainer module. Readers focused on implementation details and quick adaptation strategies will find value in the code and the reported numbers. It is less suited for those seeking major methodological breakthroughs. It deserves a serious referee. The work is grounded in existing techniques but the open implementation and the concrete validation on cross-domain transfers make it worth a closer look by the community. I would recommend sending it to peer review so that reviewers can check the code, ask for more experiments if needed, and help refine the claims.

Referee Report

1 major / 2 minor

Summary. The paper introduces Dante, an open-source training and fine-tuning backend integrated with the Dafne federated medical image segmentation framework. It supports training from scratch with automatic architecture configuration, configurable layer freezing, and an extension of Low-Rank Adaptation (LoRA) to N-dimensional convolutional layers via channel-wise factorization. Validation experiments assess Gradual Unfreezing (GU) and LoRA on cross-domain MRI transfer tasks for abdominal organ segmentation and brain white matter lesion segmentation, under both full-data and few-shot regimes, reporting up to 63.6% fewer epochs to reach 85% of peak performance with GU and Dice scores up to 0.957 with LoRA.

Significance. If the empirical claims hold under broader testing, Dante would provide a practical, open-source module for parameter-efficient adaptation of segmentation models in data-scarce clinical settings. The explicit release of code at the cited GitHub repositories is a clear strength that supports reproducibility and community adoption within the Dafne ecosystem.

major comments (1)

[Abstract] Abstract: The final sentence asserts that 'These results validate Dante as a domain-agnostic fine-tuning module for medical image segmentation in real clinical deployment conditions.' This claim is not supported by the reported experiments, which are confined to two MRI-specific tasks (abdominal organs and brain white-matter lesions) using controlled cross-domain splits on fixed architectures; no results are shown for CT/PET modalities, scanner/protocol variation, or the federated multi-site training and privacy constraints that define the Dafne framework.

minor comments (2)

[Abstract] The abstract and results sections do not report dataset sizes, exact baseline implementations, or statistical significance tests for the reported Dice and epoch reductions, making it difficult to assess the magnitude and reliability of the gains.
Notation for the LoRA extension to N-D convolutions (channel-wise factorization) is introduced without an explicit equation or pseudocode in the methods description, which would aid reproducibility.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive feedback. We agree that the abstract's concluding claim overreaches the scope of the presented experiments and will revise it accordingly.

read point-by-point responses

Referee: [Abstract] Abstract: The final sentence asserts that 'These results validate Dante as a domain-agnostic fine-tuning module for medical image segmentation in real clinical deployment conditions.' This claim is not supported by the reported experiments, which are confined to two MRI-specific tasks (abdominal organs and brain white-matter lesions) using controlled cross-domain splits on fixed architectures; no results are shown for CT/PET modalities, scanner/protocol variation, or the federated multi-site training and privacy constraints that define the Dafne framework.

Authors: We acknowledge that this sentence overstates the generalizability of the results. The validation is limited to MRI cross-domain transfer tasks on two segmentation problems and does not include other modalities, scanner variations, or explicit federated multi-site experiments. We will revise the abstract to state that the results demonstrate the effectiveness of Dante for parameter-efficient adaptation in cross-domain MRI segmentation, while noting that broader validation across modalities and federated settings remains future work. The revised sentence will read: 'These results support the use of Dante for efficient fine-tuning in cross-domain MRI medical image segmentation tasks.' This change will be made in the resubmitted version. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical validation relies on new experiments

full rationale

The paper describes the Dante module and its features (GU, LoRA for N-D conv layers) then reports direct empirical results on cross-domain MRI transfers for abdominal organs and brain lesions under full-data and few-shot regimes. Performance metrics (epoch reduction up to 63.6%, Dice up to 0.957) are obtained from fresh training runs and comparisons against baselines on held-out splits; no equations, fitted parameters renamed as predictions, self-definitional constructs, or load-bearing self-citations appear in the derivation. The validation chain is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

This is an engineering and software tool paper; no free parameters, mathematical axioms, or invented scientific entities are introduced beyond standard deep learning components.

pith-pipeline@v0.9.0 · 5598 in / 1099 out tokens · 61164 ms · 2026-05-07T12:54:35.304904+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Reference graph

Works this paper leans on

22 extracted references · 21 canonical work pages · 2 internal anchors

[1]

Generalist models in medical image segmentation: A survey and performance comparison with task-specific approaches

Moglia A, Leccardi M, Cavicchioli M, Maccarini A, Marcon M, Mainardi L, et al. Generalist models in medical image segmentation: A survey and performance comparison with task-specific approaches. Information Fusion. 2026 Mar;127:103709. doi:10.1016/j.inffus.2025.103709

work page doi:10.1016/j.inffus.2025.103709 2026
[2]

A narrative review of foundation models for medical image segmentation: zero-shot performance evaluation on diverse modalities

Noh S, Lee BD. A narrative review of foundation models for medical image segmentation: zero-shot performance evaluation on diverse modalities. Quant Imaging Med Surg. 2025 Jun;15(6):5825–58. doi:10.21037/qims-2024-2826

work page doi:10.21037/qims-2024-2826 2025
[3]

Deep Learning-Based 3D and 2D Approaches for Skeletal Muscle Segmentation on Low-Dose CT Images

Timpano G, Veltri P, Vizza P, Cascini GL, Manti F. Deep Learning-Based 3D and 2D Approaches for Skeletal Muscle Segmentation on Low-Dose CT Images. J Digit Imaging Inform med. 2025 Aug 27. doi:10.1007/s10278-025-01646-9

work page doi:10.1007/s10278-025-01646-9 2025
[4]

, author He, Y

Ma J, He Y, Li F, Han L, You C, Wang B. Segment anything in medical images. Nat Commun. 2024 Jan 22;15(1):654. doi:10.1038/s41467-024-44824-z

work page doi:10.1038/s41467-024-44824-z 2024
[5]

TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images

Wasserthal J, Breit HC, Meyer MT, Pradella M, Hinck D, Sauter AW, et al. TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images. Radiology: Artificial Intelligence. 2023 Sep 1;5(5):e230024. doi:10.1148/ryai.230024

work page doi:10.1148/ryai.230024 2023
[6]

Domain Adaptation for Medical Image Analysis: A Survey

Guan H, Liu M. Domain Adaptation for Medical Image Analysis: A Survey. IEEE Trans Biomed Eng. 2022 Mar;69(3):1173–85. doi:10.1109/TBME.2021.3117407

work page doi:10.1109/tbme.2021.3117407 2022
[7]

Transfer learning in medical image segmentation: New insights from analysis of the dynamics of model parameters and learned representations

Karimi D, Warfield SK, Gholipour A. Transfer learning in medical image segmentation: New insights from analysis of the dynamics of model parameters and learned representations. Artificial Intelligence in Medicine. 2021 Jun;116:102078. doi:10.1016/j.artmed.2021.102078

work page doi:10.1016/j.artmed.2021.102078 2021
[8]

Deep Anatomical Federated Network (Dafne): An Open Client-Server Framework for Continuous, Collaborative Improvement of Deep Learning–based Medical Image Segmentation

Santini F, Wasserthal J, Agosti A, Deligianni X, Keene KR, Kan HE, et al. Deep Anatomical Federated Network (Dafne): An Open Client-Server Framework for Continuous, Collaborative Improvement of Deep Learning–based Medical Image Segmentation. Radiology: Artificial Intelligence. 2025 May 1;7(3):e240097. doi:10.1148/ryai.240097

work page doi:10.1148/ryai.240097 2025
[9]

LoRA: Low-Rank Adaptation of Large Language Models

Hu EJ, Shen Y, Wallis P, Allen-Zhu Z, Li Y, Wang S, et al. LoRA: Low-Rank Adaptation of Large Language Models [Internet]. arXiv; 2021 [cited 2026 Mar 10]. Available from: http://arxiv.org/abs/2106.09685 doi:10.48550/arXiv.2106.09685

work page internal anchor Pith review doi:10.48550/arxiv.2106.09685 2021
[10]

arXiv preprint arXiv:2304.13785 (2023)

Zhang K, Liu D. Customized Segment Anything Model for Medical Image Segmentation [Internet]. arXiv; 2023 [cited 2026 Mar 10]. Available from: https://arxiv.org/abs/2304.13785 doi:10.48550/ARXIV.2304.13785

work page doi:10.48550/arxiv.2304.13785 2023
[11]

Convolution meets LoRA: Parameter efficient finetuning for Segment Anything Model.arXiv preprint arXiv:2401.17868, 2024

Zhong Z, Tang Z, He T, Fang H, Yuan C. Convolution Meets LoRA: Parameter Efficient Finetuning for Segment Anything Model [Internet]. arXiv; 2024 [cited 2026 Mar 10]. Available from: https://arxiv.org/abs/2401.17868 doi:10.48550/ARXIV.2401.17868

work page doi:10.48550/arxiv.2401.17868 2024
[12]

Parameter-efficient fine-tuning for medical image analysis: The missed opportunity,

Dutt R, Ericsson L, Sanchez P, Tsaftaris SA, Hospedales T. Parameter-Efficient Fine-Tuning for Medical Image Analysis: The Missed Opportunity [Internet]. arXiv; 2023 [cited 2026 Mar 10]. Available from: https://arxiv.org/abs/2305.08252 doi:10.48550/ARXIV.2305.08252

work page doi:10.48550/arxiv.2305.08252 2023
[13]

Towards Foundation Models and Few-Shot Parameter-Efficient Fine-Tuning for Volumetric Organ Segmentation

Silva-Rodríguez J, Dolz J, Ben Ayed I. Towards Foundation Models and Few-Shot Parameter-Efficient Fine-Tuning for Volumetric Organ Segmentation. Medical Image Analysis. 2025 Jul;103:103596. doi:10.1016/j.media.2025.103596

work page doi:10.1016/j.media.2025.103596 2025
[14]

MRISegmenter: A Fully Accurate and Robust Automated Multiorgan and Structure Segmentation Tool for T1-weighted Abdominal MRI

Zhuang Y, Mathai TS, Mukherjee P, Khoury B, Kim B, Hou B, et al. MRISegmenter: A Fully Accurate and Robust Automated Multiorgan and Structure Segmentation Tool for T1-weighted Abdominal MRI. Radiology. 2025 Apr 1;315(1):e241979. doi:10.1148/radiol.241979

work page doi:10.1148/radiol.241979 2025
[15]

arXiv preprint arXiv:1411.1792 , year=

Yosinski J, Clune J, Bengio Y, Lipson H. How transferable are features in deep neural networks? [Internet]. 2014. doi:10.48550/ARXIV.1411.1792

work page doi:10.48550/arxiv.1411.1792 2014
[16]

nnU-Net: Self-adapting Framework for U-Net-Based Medical Image Segmentation [Internet]

Isensee F, Petersen J, Klein A, Zimmerer D, Jaeger PF, Kohl S, et al. nnU-Net: Self-adapting Framework for U-Net-Based Medical Image Segmentation [Internet]. arXiv; 2018 [cited 2026 Mar 15]. Available from: http://arxiv.org/abs/1809.10486 doi:10.48550/arXiv.1809.10486

work page doi:10.48550/arxiv.1809.10486 2018
[17]

MONAI: An open-source framework for deep learning in healthcare

Cardoso MJ, Li W, Brown R, Ma N, Kerfoot E, Wang Y, et al. MONAI: An open-source framework for deep learning in healthcare [Internet]. arXiv; 2022 [cited 2026 Mar 19]. Available from: http://arxiv.org/abs/2211.02701 doi:10.48550/arXiv.2211.02701

work page internal anchor Pith review doi:10.48550/arxiv.2211.02701 2022
[18]

Three mechanisms of weight decay regularization, 2018

Zhang G, Wang C, Xu B, Grosse R. Three Mechanisms of Weight Decay Regularization [Internet]. arXiv; 2018 [cited 2026 May 5]. Available from: http://arxiv.org/abs/1810.12281 doi:10.48550/arXiv.1810.12281

work page doi:10.48550/arxiv.1810.12281 2018
[19]

AMOS: A Large-Scale Abdominal Multi- Organ Benchmark for Versatile Medical Image Segmentation

Ji Y, Bai H, GE C, Yang J, Zhu Y, Zhang R, et al. AMOS: A Large-Scale Abdominal Multi- Organ Benchmark for Versatile Medical Image Segmentation. In: Koyejo S, Mohamed S, Agarwal A, Belgrave D, Cho K, Oh A, editors. Advances in Neural Information Processing Systems [Internet]. Curran Associates, Inc.; 2022. p. 36722–32. Available from: https://proceedings....

2022
[20]

CHAOS - Combined (CT-MR) Healthy Abdominal Organ Segmentation Challenge Data [Internet]

Kavur AE, Selver MA, Dicle O, Barış M, Gezer NS. CHAOS - Combined (CT-MR) Healthy Abdominal Organ Segmentation Challenge Data [Internet]. Zenodo; 2019 [cited 2026 Mar 12]. Available from: https://zenodo.org/record/3362844 doi:10.5281/ZENODO.3362844

work page doi:10.5281/zenodo.3362844 2019
[21]

MSLesSeg: baseline and benchmarking of a new Multiple Sclerosis Lesion Segmentation dataset.Scientific Data, 12(1):920, May 2025

Guarnera F, Rondinella A, Crispino E, Russo G, Di Lorenzo C, Maimone D, et al. MSLesSeg: baseline and benchmarking of a new Multiple Sclerosis Lesion Segmentation dataset. Sci Data. 2025 May 31;12(1):920. doi:10.1038/s41597-025-05250-y

work page doi:10.1038/s41597-025-05250-y 2025
[22]

Longitudinal multiple sclerosis lesion segmentation data resource

Carass A, Roy S, Jog A, Cuzzocreo JL, Magrath E, Gherman A, et al. Longitudinal multiple sclerosis lesion segmentation data resource. Data in Brief. 2017 Jun;12:346–50. doi:10.1016/j.dib.2017.04.004 Supplementary materials Figure S1. Validation loss and DICE curves in the AMOS subset training. The loss and DICE curves on the validation dataset indicate th...

work page doi:10.1016/j.dib.2017.04.004 2017