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

arXiv:2605.03602 · eess.IV · Submitted 2026-05-05 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: We're looking at the paper: "Dante: An Open Source Model Pre-Training and Fine-Tuning Tool for the Dafne Federated Framework for Medical Image Segmentation".

Dev: Here's the gist of the paper. Automated segmentation techniques in the medical field have exhibited human-like performance on real clinical cases, significantly reducing the reliance on human interpretation <ref:2605.03602#pg2> Deep learning architectures, especially those based on the U-Net encoder-decoder paradigm and its variants like Vision Transformers (ViT) and hybrid models such as UNetR and Swin-Unet, as well as large-scale foundation models (1,2), have shown state-of-the-art performance in public benchmarks <ref:2605.03602#pg2> A significant challenge to the practical implementation of these models is domain shift: the systematic discrepancy between the data distribution encountered during training and the one observed during inference <ref:2605.03602#pg2> This issue represents a significant barrier to the widespread implementation of high-performance segmentation models in diverse clinical environments <ref:2605.03602#pg2>.

Rosa: Transfer learning from a pre-trained model helps mitigate this constraint <ref:2605.03602#pg2> Transfer learning from pre-trained models is the standard strategy for addressing data sparsity and domain shift in medical image segmentation <ref:2605.03602#pg2> The main theoretical basis is the concept that early convolutional layers encode general, transferable features, while deeper layers encode task-specific representations <ref:2605.03602#pg2> This layer specialization has led to studies on CNNs showing that fine-tuned CNNs consistently outperform models trained from scratch in medical imaging tasks, particularly with sparse target domain annotations <ref:2605.03602#pg2> Additionally, the depth of adaptation should correspond to the extent of domain shift (10) and encoder filters change minimally during fine-tuning in cross-modal and crossorgan transfer scenarios (7) <ref:2605.03602#pg2>.

Dev: Dante Architecture and Fine-Tuning Strategies Dante is an open-source module that serves as an automated training and fine-tuning backend within the Dafne ecosystem <ref:2605.03602#pg2> Alongside training from scratch with automatic architecture adaptation to the input data, it features two complementary selective fine-tuning strategies for segmentation models: Layers Freezing (with also Gradual Unfreezing strategy) and LoRA (Low-Rank Adaptation) applied to convolutional and transposed-convolutional layers of segmentation models <ref:2605.03602#pg2> Dante provides three base architectures for developing segmentation models from scratch: 2D U-Net, 3D U-Net, and DynUNet <ref:2605.03602#pg2> Fine-tuning strategies like Layers Freezing and LoRA are available for all three architectures, allowing users to adapt any pre-trained model to new local data <ref:2605.03602#pg2>.

Rosa: Data Preparation and Model Configuration All three architectures utilize a shared data preparation pipeline <ref:2605.03602#pg2> Input data is retrieved from numpy bundle files in a Dafne-specific format (npz bundle files) containing the image volume, segmentation masks, and original voxel resolution <ref:2605.03602#pg2> For DynUNet network hyperparameters and batch size are automatically determined based on a dataset fingerprint that summarizes the median voxel spacing and median image shape of the training set <ref:2605.03602#pg2> Augmentation operations, manually selectable by the user, include random inversions, rotations, zooms, and the addition of Gaussian noise <ref:2605.03602#pg2> For 3D model training, individual volumes are divided into patches (automatically or manually defined) using a 3:1 positive-to-negative labeling ratio <ref:2605.03602#pg2>.

Paper summary: Rosa: To unpack that a bit more, the paper says the core claim is about solving that data shift problem you mentioned earlier. They’re talking about how transfer learning, using pre-trained models, can help fix the issue where the training data looks totally different from what you see when you try to use it in a new clinic. Dev Exactly. The abstract points out that transfer learning from pre-trained models is the standard way to deal with data sparsity and that domain shift problem in medical image segmentation >

Taro: I wonder how they are actually doing this adaptation without needing massive amounts of new labeled data for every single setup? Like, what's the trick for making it work locally? Rosa That’s a good question, Taro. They focus on selective and parameter-efficient fine-tuning strategies. Dante offers two main paths: you can train a model from scratch or you can adapt an existing one using only a limited set of annotated local data >

Dev: And the key idea there is that early convolutional layers encode general features that are transferable, while the deeper layers handle the specific task, which means fine-tuned models consistently outperform models trained from scratch when you only have sparse target domain annotations >

Taro: So it’s not just about picking a model; it’s about how deep you let it get into adapting to the new data distribution. The paper mentions that the depth of adaptation should match how much the domain shift is, and encoder filters change very little during fine-tuning in cross-modal or crossorgan transfer scenarios >

Rosa: That sounds like they are trying to find a sweet spot for adaptation—not too much, not too little—to make sure you get good performance without completely forgetting what the model learned originally >

Conclusion: Rosa: So wrapping up this paper on "Dante: An Open Source Model Pre-Training and Fine-Tuning Tool for the Dafne Federated Framework for Medical Image Segmentation," what does this actually mean in practice? Dev It means Dante is positioning itself as a really valuable tool within the Dafne ecosystem, especially for research settings where data might be scarce. Rosa Exactly. They say it could simplify the pre-training process so that it can happen right there in a clinical research setting, which is great because custom software implementation isn't required >

Dev: And they highlight that this approach could benefit from a simpler user interface, and you don't need to write all that custom code yourself > Taro I guess the implication for people who just listen is that they can start leveraging existing, powerful models much faster when dealing with localized data constraints >

Rosa: Right. The main thing is this tool provides a configurable, ready-to-deploy fine-tuning backend for federated medical image segmentation systems > Dev So it’s less about creating a brand new segmentation architecture and more about giving existing architectures a smart way to adapt them locally to new problems >

Taro: But the paper does flag that there’s still work to do. They mention that the generalizability of these performance trends across different anatomical regions, imaging modalities, and clinical sites needs further investigation > Rosa That makes sense. The comparisons they made in this work are pretty focused only on Dante's fine-tuning strategies and don't include direct numerical comparisons with other parameter efficient fine-tuning frameworks in the literature > Dev So they suggest that a wider exploration of more architectures and domain shifts is needed before we can really say we have definitive guidance on which specific settings to use >

Magna Graecia University · University of Basel

eess.IV

Submitted: 2026-05-05

Updated: 2026-11-16

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 79/100

The gist: Automated segmentation techniques in the medical field have exhibited human-like performance on real clinical cases…

Key concepts

Domain Shift
This is the problem where a model trained on one type of medical data (e.g., images from one hospital) performs poorly when applied to a different, unseen dataset (e.g., images from another hospital). The paper addresses this by using transfer learning to bridge this gap.
Transfer Learning
This technique uses knowledge gained from solving one problem (like segmenting organs in one set of images) and applies it to a related but different problem. In medical imaging, it means using features learned from massive datasets to speed up training on smaller, specific clinical datasets.
LoRA (Low-Rank Adaptation)
LoRA is a parameter-efficient fine-tuning strategy that modifies only a small subset of the model's weights instead of retraining the entire model. This makes adaptation much faster and requires less computational power when fine-tuning models for specific tasks like medical segmentation.
Dante Architecture
Dante is an open-source backend module that automates the training and fine-tuning process for segmentation models. It offers different base architectures (2D U-Net, 3D U-Net, DynUNet) and provides selective fine-tuning methods like Layer Freezing and LoRA to adapt pre-trained models.

Terminology

Summary

Automated segmentation techniques in the medical field have exhibited human-like performance on real clinical cases, significantly reducing the reliance on human interpretation <ref:2605. p. 0

Where the paper says it

Introduction In recent years, automatic segmentation techniques in the medical field have exhibited human-like performance on real clinical cases, significantly reducing the reliance on human interpretation. Automated segmentation of anatomical structures from medical images is an important step for quantitative clinical analysis to support volumetric organ assessment, surgical planning, radiotherapy contouring, and the extraction of imaging biomarkers for longitudinal follow-up.

03602#pg2> Deep learning architectures, especially those based on the U-Net encoder-decoder paradigm and its variants like Vision Transformers (ViT) and hybrid models such as UNetR and Swin-Unet, as well as large-scale foundation models (1,2), have shown state-of-the-art performance in public benchmarks <ref:2605. p. 0
Where the paper says it

Automated segmentation of anatomical structures from medical images is an important step for quantitative clinical analysis to support volumetric organ assessment, surgical planning, radiotherapy contouring, and the extraction of imaging biomarkers for longitudinal follow-up. Deep learning architectures, especially those based on the U-Net encoder-decoder paradigm and its variants like Vision Transformers (ViT) and hybrid models such as UNetR and Swin-Unet, as well as large-scale foundation models (1,2), have shown state-of-the-art performance in public benchmarks, achieving Dice similarity coefficients that approach inter-expert agreement for various anatomical structures in CT and MRI modalities (3).

03602#pg2> A significant challenge to the practical implementation of these models is domain shift: the systematic discrepancy between the data distribution encountered during training and the one observed during inference <ref:2605. p. 0
Where the paper says it

Acquiring and validating annotations from expert radiologists can be costly and impractical, which hinders the efficacy of ad-hoc models trained on local data. A significant challenge to the practical implementation of these models is domain shift: the systematic discrepancy between the data distribution encountered during training and the one observed during inference.

03602#pg2> This issue represents a significant barrier to the widespread implementation of high-performance segmentation models in diverse clinical environments <ref:2605. p. 0
Where the paper says it

Generally, a model pre-trained on an institutional dataset, particularly for MRI images (which are highly susceptible to inter-patient variability and inter-operator differences in segmentation), is unlikely to maintain its performance when applied to data from a different site or protocol without some form of adaptation (6,7). This issue represents a significant barrier to the widespread implementation of high-performance segmentation models in diverse clinical environments.

03602#pg2>.

Introduction to Transfer Learning

Transfer learning from a pre-trained model helps mitigate this constraint <ref:2605.03602#pg2> Transfer learning from pre-trained models is the standard strategy for addressing data sparsity and domain shift in medical image segmentation <ref:2605. p. 1

Where the paper says it

A detailed description of the experimental design and specific scenarios can be found in Section 3. Transfer learning from pre-trained models is the standard strategy for addressing data sparsity and domain shift in medical image segmentation.

03602#pg2> The main theoretical basis is the concept that early convolutional layers encode general, transferable features, while deeper layers encode task-specific representations <ref:2605. p. 1
Where the paper says it

Transfer learning from pre-trained models is the standard strategy for addressing data sparsity and domain shift in medical image segmentation. The main theoretical basis is the concept that early convolutional layers encode general, transferable features, while deeper layers encode task-specific representations (15).

03602#pg2> This layer specialization has led to studies on CNNs showing that fine-tuned CNNs consistently outperform models trained from scratch in medical imaging tasks, particularly with sparse target domain annotations <ref:2605. p. 1
Where the paper says it

The main theoretical basis is the concept that early convolutional layers encode general, transferable features, while deeper layers encode task-specific representations (15). This layer specialization has led to studies on CNNs showing that fine-tuned CNNs consistently outperform models trained from scratch in medical imaging tasks, particularly with sparse target domain annotations.

03602#pg2> Additionally, the depth of adaptation should correspond to the extent of domain shift (10) and encoder filters change minimally during fine-tuning in cross-modal and crossorgan transfer scenarios (7) <ref:2605. p. 1
Where the paper says it

This layer specialization has led to studies on CNNs showing that fine-tuned CNNs consistently outperform models trained from scratch in medical imaging tasks, particularly with sparse target domain annotations. Additionally, the depth of adaptation should correspond to the extent of domain shift (10), and encoder filters change minimally during fine-tuning in cross-modal and crossorgan transfer scenarios (7).

03602#pg2>.

Dante Architecture and Fine-Tuning Strategies

Dante is an open-source module that serves as an automated training and fine-tuning backend within the Dafne ecosystem <ref:2605.03602#pg2> Alongside training from scratch with automatic architecture adaptation to the input data, it features two complementary selective fine-tuning strategies for segmentation models: Layers Freezing (with also Gradual Unfreezing strategy) and LoRA (Low-Rank Adaptation) applied to convolutional and transposed-convolutional layers of segmentation models <ref:2605. p. 1

Where the paper says it

This paper introduces DAfNe TrainEr (Dante), an open-source module that serves as an automated training and fine-tuning backend within the Dafne ecosystem. Alongside training from scratch with automatic architecture adaptation to the input data, it features two complementary selective fine-tuning strategies for segmentation models: Layers Freezing (with also Gradual Unfreezing strategy) and LoRA (Low-Rank Adaptation) applied to convolutional and transposed-convolutional layers of segmentation models.

03602#pg2> Dante provides three base architectures for developing segmentation models from scratch: 2D U-Net, 3D U-Net, and DynUNet <ref:2605. p. 2
Where the paper says it

The adaptation process should focus on updating the most relevant parameters for the local data while preserving the existing knowledge, ensuring that the model's prior learning is retained and not discarded. Dante provides three base architectures for developing segmentation models from scratch: 2D U-Net, 3D U-Net, and DynUNet.

03602#pg2> Fine-tuning strategies like Layers Freezing and LoRA are available for all three architectures, allowing users to adapt any pre-trained model to new local data <ref:2605. p. 2
Where the paper says it

The DynUNet architecture is based on the automatic configuration of training parameters and network, following the self-configuration principle of nnU-Net (16), enabling it to adjust to the specific voxel spacing, image size, and modality of the target dataset without manual hyperparameter tuning. Furthermore, fine-tuning strategies like Layers Freezing and LoRA are available for all three architectures, allowing users to adapt any pre-trained model to new local data, regardless of the training mode originally used.

03602#pg2>.

Data Preparation and Model Configuration

All three architectures utilize a shared data preparation pipeline <ref:2605.03602#pg2> Input data is retrieved from numpy bundle files in a Dafne-specific format (npz bundle files) containing the image volume, segmentation masks, and original voxel resolution <ref:2605. p. 2

Where the paper says it

Data preparation All three architectures utilize a shared data preparation pipeline. Input data is retrieved from numpy bundle files in a Dafne-specific format (npz bundle files) containing the image volume, segmentation masks, and original voxel resolution (Figure 1).

03602#pg2> For DynUNet network hyperparameters and batch size are automatically determined based on a dataset fingerprint that summarizes the median voxel spacing and median image shape of the training set <ref:2605. p. 2
Where the paper says it

Intensity normalization is applied to non-zero voxels on a channel-by-channel basis, preventing background regions from influencing the normalization statistics. For DynUNet network hyperparameters and batch size are automatically determined based on a dataset fingerprint that summarizes the median voxel spacing and median image shape of the training set.

03602#pg2> Augmentation operations, manually selectable by the user, include random inversions, rotations, zooms, and the addition of Gaussian noise <ref:2605. p. 2
Where the paper says it

In contrast, for 3D and 2D U-Nets, some parameters such as the patch size (for the sliding window training and inference (17)) are fixed, while others, such as the batch size and network hyperparameters, can be manually defined by the user (see Table 1 for more details). Augmentation operations, manually selectable by the user, include random inversions, rotations, zooms, and the addition of Gaussian noise.

03602#pg2> For 3D model training, individual volumes are divided into patches (automatically or manually defined) using a 3:1 positive-to-negative labeling ratio <ref:2605. p. 3
Where the paper says it

volumes are divided into patches (automatically or manually defined) using a 3:1 positive-to-negative labeling ratio, ensuring that each batch contains a sufficient proportion of annotated foreground regions. For the 2D U-Net, patch extraction is replaced with slice-level selection, retaining only slices that contain at least one positive annotation along with their immediate surroundings (Table 1).

03602#pg2>.

Performance of Adaptation Strategies

The performance of the system was tested under the following conditions: training from scratch of all three architectures on the CHAOS dataset, and multi-organ transfer learning by using the various combinations of GU and LoRA strategies starting from base models pretrained on AMOS, CHAOS, NIH, and MSLesSeg datasets <ref:2605.03602#pg2> The gradual unblocking strategy emerged as the most efficient approach for quick implementation, hitting the 85% performance threshold in just 16 epochs and resulting in a 63. p. 7

Where the paper says it

8726 in only 64 epochs, indicating that the small CHAOS dataset did not provide enough morphological diversity to effectively train the network without pre-training. The gradual unblocking strategy emerged as the most efficient approach for quick implementation, hitting the 85% performance threshold in just 16 epochs and resulting in a 63.

6% reduction in training time compared to the baseline model <ref:2605. p. 7
Where the paper says it

The gradual unblocking strategy emerged as the most efficient approach for quick implementation, hitting the 85% performance threshold in just 16 epochs and resulting in a 63. 6% reduction in training time compared to the baseline model.

03602#pg2> When pre-training on the NIH dataset, all fine-tuning strategies produced notably higher DSC values: LoRA (r = 32) achieved 0. p. 8
Where the paper says it

8864, while GU reached 85% peak performance in just 6 epochs, compared to 41 epochs for the baseline model (Table 4). When pre-training on the NIH dataset, all fine-tuning strategies produced notably higher DSC values: LoRA (r = 32) achieved 0.

957, and GU reached 0. p. 8
Where the paper says it

When pre-training on the NIH dataset, all fine-tuning strategies produced notably higher DSC values: LoRA (r = 32) achieved 0. 957, and GU reached 0.

956, while the baseline model, trained on 10 patients from AMOS dataset, reached 0. p. 8
Where the paper says it

957, and GU reached 0. 956, while the baseline model, trained on 10 patients from AMOS dataset, reached 0.

798 <ref:2605.03602#pg2>.

Key Findings on Strategy Efficacy

When pre-training on the NIH dataset and fine-tuning it on the AMOS dataset, while ensuring that labels from both datasets remain mutually exclusive, the results indicate that GU achieves an exceptional balance between accuracy and convergence speed in the new AMOS domain, with a DSC of 0.9318 <ref:2605.03602#pg2> This performance surpasses that of the LoRA fine-tuning, which produced DSC scores of 0. p. 8

Where the paper says it

By pretraining the model on the NIH dataset and fine-tuning it on the AMOS dataset, while ensuring that labels from both datasets remain mutually exclusive, the results indicate that GU achieves an exceptional balance between accuracy and convergence speed in the new AMOS domain, with a DSC of 0. This performance surpasses that of the LoRA fine-tuning, which produced DSC scores of 0.

8860, 0.8990, and 0. p. 8
Where the paper says it

This performance surpasses that of the LoRA fine-tuning, which produced DSC scores of 0. 8990, and 0.

9058 for r = 8, 16, and 32 respectively <ref:2605. p. 8
Where the paper says it

8990, and 0. 9058 for r = 8, 16, and 32, respectively.

03602#pg2> The organ-specific analysis reveals varied learning dynamics: the spleen benefits from nearinstantaneous feature transfer during the first four epochs, while the liver requires a much longer adaptation phase <ref:2605. p. 9
Where the paper says it

9058) comes from the greater parametric freedom that GU provides during the initial finetuning phases, which helps restructure the decoder to manage mutually exclusive classes. Additionally, the organ-specific analysis reveals varied learning dynamics: the spleen benefits from nearinstantaneous feature transfer during the first four epochs, while the liver requires a much longer adaptation phase, spanning epochs 58 to 133.

03602#pg2>.

Conclusion and Limitations

Dante places itself as a potentially valuable tool in the Dafne ecosystem, for the simplified pretraining of models which can be achieved in a clinical research setting, where data is potentially scarce, and which might benefit from a simplified user interface without the requirement of custom software implementation <ref:2605.03602#pg2> The generalizability of the observed performance trends to other anatomical regions, imaging modalities, and clinical sites requires further investigation <ref:2605. p. 0

Where the paper says it

Automated segmentation of anatomical structures from medical images is an important step for quantitative clinical analysis to support volumetric organ assessment, surgical planning, radiotherapy contouring, and the extraction of imaging biomarkers for longitudinal follow-up. Deep learning architectures, especially those based on the U-Net encoder-decoder paradigm and its variants like Vision Transformers (ViT) and hybrid models such as UNetR and Swin-Unet, as well as large-scale foundation models (1,2), have shown state-of-the-art performance in public benchmarks, achieving Dice similarity coefficients that approach inter-expert agreement for various anatomical structures in CT and MRI modalities (3).

03602#pg2> The comparisons in this work are limited to Dante's fine-tuning strategies and do not include direct numerical comparisons with other PEFT frameworks in the literature <ref:2605.03602#pg2> A more extensive exploration of a wider range of architectures and domain shifts is necessary to provide definitive guidance on rank selection <ref:2605.03602#pg2>.

Summary of Experimental Scenarios

The experimental validation utilized five open-source MRI datasets, divided into two anatomical domains: abdominal MRI datasets and brain MRI datasets <ref:2605.03602#pg2> Five transfer-learning scenarios were established based on the pre-training and fine-tuning combinations across these diverse sources <ref:2605. p. 5

Where the paper says it

From these datasets, five transfer-learning scenarios were established. In the first abdominal scenario, pre-training was conducted on AMOS, followed by fine-tuning on CHAOS, with evaluation limited to the four structures consistently annotated in both datasets: liver, spleen, right kidney, and left kidney.

03602#pg2> These scenarios included multi-organ transfer, where models were pre-trained on one dataset and fine-tuned on another, as well as adaptation for multiple sclerosis lesion segmentation <ref:2605. p. 5
Where the paper says it

The fourth scenario involved pre-training on NIH using labels that differed from the AMOS target labels to evaluate transfer learning across different segmentation targets. Finally, in the brain MRI scenario, pre-training was performed on MSLesSeg, followed by fine-tuning on ISBI 2015 to assess adaptation for multiple sclerosis lesion segmentation in a context of extreme data scarcity.

03602#pg2>. The study demonstrated that the advantages of pre-trained representations extend beyond multi-organ segmentation tasks <ref:2605. p. 9
Where the paper says it

In both cases, GU and LoRA outperformed the baseline without any task-specific configuration, indicating that the module can be deployed as-is in federated nodes operating on different anatomical targets and imaging protocols, without the need for expert tuning for each new application. In this case, results demonstrate that the advantages of pre-trained representations extend beyond multi-organ segmentation tasks.

03602#pg2>.

Evaluation Metrics and Convergence Speed

Trends, stability, and convergence speed by measuring the average Dice Similarity Coefficient (DSC) across all segmented labels, considering the limited validation set, were analyzed <ref:2605.03602#pg2> DSC after completion of the training was considered as the primary endpoint <ref:2605. p. 6

Where the paper says it

Evaluation metrics and data analysis Trends, stability, and convergence speed by measuring the average Dice Similarity Coefficient (DSC) across all segmented labels, considering the limited validation set, were analyzed: the primary goal is to validate Dante and assess how its fine-tuning strategies can mitigate the performance penalty caused by domain shift. DSC after completion of the training was considered as the primary endpoint.

03602#pg2> For transfer learning, the number of epochs needed to reach the peak DSC performance, and the number of epochs needed to reach 85% of the maximum performance were considered as indicators of convergence speed <ref:2605. p. 6
Where the paper says it

DSC after completion of the training was considered as the primary endpoint. For transfer learning, the number of epochs needed to reach the peak DSC performance, and the number of epochs needed to reach 85% of the maximum performance were considered as indicators of convergence speed, to compare the various fine tuning algorithms.

03602#pg2>. Table 4 illustrates how parameter-efficient fine-tuning, combined with extensive anatomical pre-training, enhances high-accuracy segmentation while requiring minimal local data <ref:2605. p. 2
Where the paper says it

Furthermore, fine-tuning strategies like Layers Freezing and LoRA are available for all three architectures, allowing users to adapt any pre-trained model to new local data, regardless of the training mode originally used. Figure 1 illustrates the overall architecture of Dante and its two operational modes.

03602#pg2>. The results indicate that the NIH pretraining condition leads to significantly smoother and faster convergence curves than CHAOS pretraining <ref:2605. p. 8
Where the paper says it

Validation and dice loss curves demonstrate that the NIH pretraining condition leads to significantly smoother and faster convergence curves than CHAOS pretraining, which aligns with the higher final DSC values reported earlier (see Figure S1 in supplementary materials for more details). By pretraining the model on the NIH dataset and fine-tuning it on the AMOS dataset, while ensuring that labels from both datasets remain mutually exclusive, the results indicate that GU achieves an exceptional balance between accuracy and convergence speed in the new AMOS domain, with a DSC of 0.

03602#pg2>.

Model Architecture Comparison

A performance gap was observed between the manually defined models of UNet-2D and UNet-3D due to the challenges associated with volumetric training in data-sparse environments <ref:2605.03602#pg2> The 3D U-Net processes sub-volumes (patches) of size (16, 96, 96), a limitation imposed by GPU memory constraints, which restricts the network's effective receptive field <ref:2605.03602#pg2> This finding underscores the advantages of using the DynUNet module within Dante for processing 3D volumes <ref:2605. p. 9

Where the paper says it

Consequently, it captures less global contextual information than the 2D approach, which analyzes entire axial slices. This finding underscores the advantages of using the DynUNet module within Dante for processing 3D volumes.

03602#pg2>. Unlike conventional 3D U-Net, which relies on manually defined parameters, DynUNet automatically optimizes the patch size based on available VRAM and batch size <ref:2605. p. 9
Where the paper says it

This finding underscores the advantages of using the DynUNet module within Dante for processing 3D volumes. Unlike conventional 3D U-Net, which relies on manually defined parameters, DynUNet automatically optimizes the patch size based on available VRAM and batch size.

03602#pg2> By fine-tuning these spatial parameters, DynUNet can capture a broader anatomical context during training than UNet-3D <ref:2605. p. 9
Where the paper says it

Unlike conventional 3D U-Net, which relies on manually defined parameters, DynUNet automatically optimizes the patch size based on available VRAM and batch size. By fine-tuning these spatial parameters, DynUNet can capture a broader anatomical context during training than UNet-3D, leading to more accurate contours and improved performance.

03602#pg2>.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Automated training of segmentation models can be achieved through Dante TrainEr, which features training from scratch with automatic architecture adaptation to the input data or fine-tuning strategies like Layers Freezing (with also Gradual Unfreezing strategy) and LoRA (Low-Rank Adaptation).

  2. The improved system can mitigate domain shift by using selective fine-tuning strategies that adapt only specific subsets of model parameters, preserving prior knowledge while allowing higher layers to adapt, as suggested by the rationale: encoder filters change minimally during fine-tuning in cross-modal and crossorgan transfer scenarios.

  3. The system can handle data scarcity effectively by leveraging pre-trained models from large datasets like AMOS or NIH for initial training, ensuring that the richer and more representative the global model distributed, the more effective the local adaptation will be, even with very few target annotations.

  4. The improved AI system can optimize network configuration automatically using DynUNet to determine hyperparameters based on a dataset fingerprint, allowing it to adjust to the specific voxel spacing, image size, and modality of the target dataset without manual hyperparameter tuning.

  5. The system can rapidly converge in data-sparse scenarios by employing Gradual Unfreezing (GU), which demonstrated the ability to converge rapidly and achieving peak performance in just one epoch for some tasks compared to baseline models requiring many epochs.

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 suggest that Dante can support the adaptation of medical image segmentation models across the evaluated anatomical and imaging domains when annotated target data are limited. 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.

Sources

Related papers