PaperlayerView exact revision on arXiv

Source-linked AI summary

Novel Methods for Catheter and Guidewire Segmentation in X-ray Fluoroscopy under a Federated Learning Setting

Chayun Kongtongvattana

arXiv:2609.06876v1cs.CVcs.AI

TL;DR

Catheter and guidewire analysis is limited by thin-structure complexity, scarce annotations, and privacy constraints on centralised training. This thesis combines a multi-domain benchmark with shape-aware losses, federated optimisation, and structure-aware synthetic video generation, reporting improved segmentation and data-efficient learning across distributed settings.

  • Problem

    Catheter and guidewire segmentation is challenged by difficult fluoroscopic imaging, limited annotations, and heterogeneous federated data distributions.

  • Method

    The thesis combines CathAction, signed-distance-map and transformer-based shape supervision, FedSSL, FedPGD, and structure-aware diffusion for synthetic fluoroscopic video.

  • Results

    The methods consistently improve segmentation across centralised and federated settings, while synthetic sequences improve performance under data scarcity.

  • Takeaways & Limitations

    The contributions support privacy-preserving collaborative catheter and guidewire analysis without centralising sensitive patient data or relying on extensive manual annotation.

  • Takeaways & Limitations

    The evaluated methods require broader validation across scanner types, institutions, patient populations, and more diverse clinical acquisition conditions.

Abstract

from arXiv · show

Endovascular procedures rely on real-time manipulation of thin instruments, catheters and guidewires, under X-ray fluoroscopy guidance, where accurate visual analysis is essential for procedural safety. Learning-based methods are constrained by structural complexity, data scarcity, and privacy regulations precluding centralised training across institutions. This thesis presents a structure-aware federated learning framework for catheter and guidewire analysis, with four contributions evaluated on real-animal and phantom data. A benchmark dataset, CathAction, is introduced for catheterisation analysis, with over 600,000 annotated frames and 40,000 segmentation masks. A shape-sensitive loss transforms masks into signed distance maps compared in a structural feature space, improving Dice coefficient by up to 2.9 points across five backbones. This is extended to federated learning with shape-sensitive loss, preserving geometric consistency under heterogeneous client data and outperforming federated averaging by up to three points in mean intersection-over-union as clients scale from four to eight. Federated learning with projected gradient descent adds adversarial optimisation, raising mean intersection-over-union by over ten points on real-animal data. Finally, a structure-aware diffusion framework synthesises catheter and guidewire video sequences, combining structural supervision with a domain-adaptive reconstruction objective, reducing Frechet video distance over a strong baseline while maintaining visual fidelity. Incorporating synthetic sequences into federated training raises the Dice score from 44 to 51 percent under data scarcity, with gains across four held-out sites. Together, these contributions advance privacy-preserving catheter and guidewire analysis, supporting collaborative training without centralising patient data or large amounts of manual annotation.

Declaration

The thesis develops structure-aware methods for catheter and guidewire segmentation under federated, privacy-sensitive clinical constraints. Its contributions span benchmark data, shape-aware supervision, robust federated optimisation, and synthetic data generation.

  • Thesis Contributions: Together, the contributions target reliable, generalisable catheter and guidewire segmentation without centralising sensitive patient data.The thesis connects structural representation, benchmark development, federated robustness, and data scarcity within privacy-preserving learning.
  • Thesis Contributions: CathAction addresses limited benchmark data by spanning phantom, animal, and human domains for endovascular analysis.It supports evaluation of structure-aware segmentation and federated learning under heterogeneous conditions.
  • Thesis Contributions: FedSSL extends structure-aware supervision to federated learning, improving segmentation consistency across institutions with Non-IID data.The framework aligns local predictions, annotations, and global model outputs.
  • Thesis Contributions: FedPGD integrates adversarial optimisation into federated learning to improve robustness, convergence stability, and generalisation across heterogeneous clients.It complements structural supervision with robustness-oriented optimisation.
  • Thesis Contributions: A structure-aware diffusion framework generates synthetic X-ray video sequences to support data-efficient training when local annotations are scarce.The generated sequences are designed to preserve instrument continuity and support centralised and federated learning.
  • Thesis Contributions: The Shape-Sensitive Loss uses signed distance maps and transformer embeddings to capture continuity in thin, elongated instruments.This addresses the limitations of pixel-wise objectives that do not explicitly model structural continuity.

1.3 Thesis Contributions

The thesis addresses limited data, structural discontinuity, and heterogeneous federated training in catheter and guidewire analysis. It combines a multi-domain benchmark with shape-aware segmentation, robust federated optimisation, and synthetic video generation.

  • 1.3 Thesis Contributions: Existing methods are constrained by scarce annotations, structure-insensitive losses, and degraded federated performance under heterogeneous data distributions.These limitations affect segmentation of thin instruments in endovascular fluoroscopic imaging.
  • 1.3 Thesis Contributions: The Shape-Sensitive Loss encodes instrument geometry with signed distance maps and vision-transformer feature similarity.It is designed to improve thin-instrument segmentation under challenging imaging conditions.
  • 1.3 Thesis Contributions: FedSSL incorporates structure-aware supervision into distributed training to improve segmentation consistency across sites.Its federated design addresses structural alignment under Non-IID data distributions.
  • 1.3 Thesis Contributions: FedPGD uses adversarial optimisation to improve robustness and convergence under heterogeneous clinical data.The thesis reports consistent advantages over FedAvg and FedProx across datasets and client configurations.
  • 1.3 Thesis Contributions: Structure-aware diffusion generates temporally coherent synthetic X-ray sequences that preserve instrument continuity for data-efficient learning.Using generated data improved downstream segmentation under limited annotations and across held-out acquisition sites.
  • 1.3 Thesis Contributions: CathAction combines phantom, animal, and human acquisitions for catheterisation segmentation, action recognition, and collision detection.The dataset contains approximately 600,000 annotated frames and 40,000 pixel-level masks.

1.5 Outline

The thesis progresses from dataset development to structure-aware segmentation, federated optimisation, and synthetic video generation for catheter and guidewire analysis under distributed clinical constraints. It targets low-contrast thin instruments, heterogeneous institutional data, and limited annotations while preserving geometric continuity.

  • Chapter 3: CathAction provides a collaborative benchmark for catheter and guidewire segmentation across real-animal, phantom, and human X-ray acquisitions.The dataset also supports catheterisation action recognition and collision detection.
  • Chapter 4: The Shape-Sensitive Loss encodes instrument geometry with signed distance maps and transformer-based feature similarity to improve thin-structure segmentation.Its purpose is to encourage geometric consistency in predicted structures.
  • Chapter 5: FedSSL promotes structural consistency under Non-IID client data, while FedPGD improves federated robustness and stability through adversarial optimisation.These methods extend the structure-aware approach to distributed clinical settings.
  • Chapter 6: A structure-aware diffusion framework generates synthetic X-ray video sequences to support centralised and federated training under limited-data conditions.The framework combines structural supervision with a domain-adaptive reconstruction objective to preserve thin-instrument geometry during generation.
  • Technical challenges: Thin catheters and guidewires are difficult to segment because they are low-contrast, sparse, and geometrically continuous structures in ambiguous fluoroscopic images.Projection overlap, low-dose noise, and cross-site variability further complicate reliable delineation.
  • Technical challenges: Pixel-wise objectives can achieve high overall accuracy while producing fragmented predictions that fail to preserve structural connectivity.This motivates structure-aware loss design for thin elongated instruments.

2.3 Federated Learning for Endovascular Fluoroscopic Imaging

Federated learning enables collaborative model training while keeping medical data local, addressing privacy constraints and heterogeneous data distributions across institutions. Its workflow alternates local optimisation with server-side aggregation, with FedAvg balancing communication cost and global alignment.

  • Centralised learning aggregates all training data, but medical imaging data often cannot be pooled because of privacy regulations, institutional policies, and ethical constraints.
  • Classical distributed training partitions data across workers and synchronises updates within shared trusted infrastructure, assuming approximately i.i.d. data.
  • Unlike classical distributed training, federated learning is designed for statistically heterogeneous, non-i.i.d. client data arising from different patient populations and acquisition conditions.
  • Federated learning keeps client datasets local while clients train privately and periodically send model updates to a central server for aggregation.
  • FedAvg alternates local optimisation with weighted global aggregation, while more local steps reduce communication but may amplify client drift under heterogeneous distributions.
  • Clinical federated optimisation must address decentralised, non-identically distributed data because local updates can drift toward client-specific optima and degrade global performance.

2.4 Adversarial Training

Adversarial training constructs bounded input perturbations and optimises models against them to improve robustness. In federated fluoroscopic segmentation, this motivation is linked to cross-site appearance shifts, while diffusion models provide a route to augment scarce heterogeneous client data.

  • Adversarial Training: Adversarial examples are visually similar inputs formed by adding small bounded perturbations that can significantly alter neural-network predictions.
  • Adversarial Training: Adversarial training optimises models to maintain correct predictions under perturbed inputs through a min–max objective.
  • Adversarial Training: Projected gradient descent is widely used as an approximate method for constructing strong adversarial examples because exact inner maximisation is intractable.
  • Federated Fluoroscopic Segmentation: Fluoroscopic models face site-specific appearance shifts and may benefit from adversarial objectives that encourage invariant representations, although robustness can trade off against segmentation accuracy.
  • Diffusion-Based Generation: Small institutional datasets and privacy restrictions motivate generative augmentation to expand local data and improve distributional coverage across heterogeneous sites.
  • Diffusion-Based Generation: Diffusion models progressively corrupt clean samples with Gaussian noise and learn a reverse denoising process that generates structured samples from noise.
  • Diffusion-Based Generation: Latent diffusion performs this process in compressed latent space, reducing computational cost while maintaining high visual fidelity for federated fluoroscopic generation.
  • Diffusion-Based Generation: Site-specific synthetic fluoroscopic images can augment underrepresented client datasets and reduce distributional heterogeneity that drives client drift.

CathAction: A Large-Scale Benchmark for Endovascular Intervention Analysis

CathAction addresses the limited scale, privacy, and task scope of existing endovascular datasets with a large multi-source benchmark. It provides annotated data and benchmarks for segmentation, action understanding, collision detection, and related procedural analysis.

  • Motivation: Existing endovascular datasets are often small, private, and restricted to narrow tasks, limiting broader procedural understanding and deep-learning development.
  • Dataset Construction: CathAction combines silicon phantom, preclinical animal, and real-world human procedures into a large-scale benchmark for endovascular intervention analysis.
  • Dataset Scale: Approximately 600,000 annotated frames support action recognition and collision detection, while 40,000 ground-truth masks support catheter and guidewire segmentation.
  • Benchmarks: CathAction benchmarks catheterisation anticipation, recognition, segmentation, collision detection, and domain adaptation, with code and data made publicly available.
  • Domain Coverage: CathAction includes phantom and real-animal data and reveals a substantial domain gap between controlled simulation and real-world settings.
  • Segmentation Labels: The dataset labels catheters and guidewires as separate classes, providing more detailed segmentation ground truth than resources that combine them.

3.4 Tasks and Benchmark Evaluation

The benchmark evaluates anticipation, recognition, segmentation, collision detection, and domain adaptation across phantom, animal, and human data. Results show transformer-based or structure-aware methods can perform well in-domain, but domain shifts and thin-instrument appearance remain substantial challenges.

  • Evaluation Setup: The evaluation covers anticipation, recognition, segmentation, collision detection, and domain adaptation, reserving human data for out-of-domain evaluation.
  • Catheterisation Anticipation: Transformer-based anticipation methods outperform CNN- and LSTM-based baselines, with AFFT exceeding RU-LSTM by 2.83%, 2.58%, and 2.86% on Accuracy, Precision, and Recall.
  • Catheterisation Recognition: TDN with ResNet101 achieves the best top-1 accuracy of 62.5% for five-class catheterisation recognition.
  • Segmentation: SegViT and other transformer-based segmentation networks outperform traditional UNet, although the improvement over competing methods is not large.
  • Segmentation Challenges: The benchmark attributes segmentation difficulty to extremely slender instruments and the distinctive characteristics of X-ray images.
  • Collision Detection: Collision detection remains difficult: tiny-object detectors perform best, but the top mean AP is only 16.91.
  • Domain Adaptation: Domain adaptation produces substantial accuracy drops, with performance declining approximately 30% relative to the recognition setup and remaining low for collision detection.
  • Human and Cross-Domain Evaluation: SSL shows strong overall performance for thin elongated structures in real human X-ray images, while SwinUNet and SSL demonstrate relatively better robustness under domain shift.

3.5 Discussion and Conclusion

This section presents CathAction as a standardised, multi-domain benchmark and positions structure-aware segmentation as a response to the thin instruments' geometric challenges. It also acknowledges gaps in coverage, generalisability, annotation quality control, and clinical relevance of current metrics.

  • CathAction dataset: CathAction provides annotated ground truth for catheter and guidewire segmentation, action understanding, and collision detection across phantom, animal, and human procedures.The dataset also supplies standardised benchmarks, public code, and metrics for reproducible evaluation.
  • CathAction dataset: The dataset addresses limited endovascular resources by supporting broader procedural understanding beyond segmentation alone.Its scope includes action recognition and collision detection, which existing resources often lack.
  • Limitations: The dataset may omit rare or outlier clinical scenarios, while current methods retain insufficient accuracy and limited generalisability to real-world settings.The authors also identify incomplete reporting of systematic quality control for collision and segmentation annotations and possible mismatch between adopted metrics and clinical needs.
  • Shape-Sensitive Loss: Shape-sensitive loss represents predicted and ground-truth masks as signed distance maps and compares learned structural embeddings to encourage geometric continuity.The method uses a vision transformer and feature-based similarity rather than relying only on pixel-wise or regional overlap objectives.
  • Shape-Sensitive Loss: Thin elongated instruments, low contrast, and background clutter can produce fragmented predictions because conventional losses do not explicitly model spatial relationships or structural continuity.These failures may remain despite locally accurate pixel predictions.

4.3 Methodology

The methodology converts segmentation masks into signed distance maps and uses transformer-derived feature similarity to make the loss sensitive to global geometry. This augments Dice loss with structural alignment information for thin, elongated instruments.

  • SDM transformation: Signed distance maps encode each pixel's signed distance to the target boundary, providing continuous geometric information beyond binary labels.Interior pixels are negative, boundary pixels are zero, and exterior pixels are positive.
  • SDM transformation: Predicted and ground-truth masks are converted into signed distance maps after contour extraction before loss computation.The pipeline begins by thresholding the network prediction to obtain a binary contour representation.
  • Structural feature extraction: A vision transformer extracts high-level features from both signed distance maps, capturing long-range geometric dependencies and global structural characteristics.This addresses the locality of convolutional operations for elongated instruments spanning large image regions.
  • Shape-sensitive loss: Cosine similarity between predicted and ground-truth SDM embeddings makes the learning objective sensitive to global geometric consistency.The resulting shape-sensitive loss measures deviation from perfect feature alignment.
  • Shape-sensitive loss: The total loss combines Dice loss with the shape-sensitive loss to balance regional segmentation accuracy and structural sensitivity.This composite objective supplies a multifaceted learning signal for segmentation.

4.4 Experiments

Experiments show that shape-aware supervision improves thin-structure segmentation across architectures and domains, while FedSSL and FedPGD provide complementary structural and robustness benefits in federated settings.

  • Shape-sensitive loss: Dice coefficient increased from 54.52% to 57.16% with TransU-Net, a gain of 2.64 percentage points.The improvement was observed across the evaluated segmentation architectures and metrics.
  • Shape-sensitive loss: Dice coefficient increased from 40.83% to 43.71% when the proposed loss was combined with TransU-Net on phantom data.The gain indicates that structural supervision remained beneficial in the controlled phantom setting.
  • Ablation and loss weighting: Equal weighting of Dice and structural objectives achieved a Dice coefficient of 57.16% for TransU-Net, while tested weighting changes produced only small score differences.Scores ranged from 56.68% to 57.16% for TransU-Net and from 54.15% to 54.71% for SwinU-Net.
  • Qualitative results: The proposed loss preserved coherent, continuous catheter and guidewire predictions, reducing fragmentation around distal tips and thin guidewire segments.These geometric improvements remained visible across animal and phantom domains despite differences in contrast and background complexity.
  • Discussion and limitation: Further validation across more diverse clinical acquisition conditions is necessary to establish broader generalisability.The reported improvements were consistent across the evaluated architectures and datasets, but the authors identify broader clinical validation as an open requirement.
  • Federated framework: FedSSL incorporates shape-sensitive supervision to improve geometric consistency across heterogeneous clients, while FedPGD uses adversarial optimisation to improve robustness under divergent client distributions.Together, the methods target structural fidelity and optimisation stability as complementary requirements for federated thin-structure segmentation.
  • Federated framework: Shape-aware supervision and robust optimisation consistently outperformed conventional federated approaches, supporting task-specific design for thin-structure segmentation.The framework couples local structural guidance with global information and adversarial optimisation tailored to heterogeneous federated data.

5.5 Conclusion

FedSSL and FedPGD provide complementary federated approaches for improving catheter and guidewire segmentation under heterogeneous clinical data. Their evaluation also exposes deployment limitations, including fixed-round training and the assumption of reliable client participation.

  • FedSSL incorporates shape-sensitive supervision to preserve geometric structure, while FedPGD uses adversarial optimisation to improve robustness and convergence stability.Both methods target thin-structure segmentation under heterogeneous data distributions.
  • Both methods outperform FedAvg and FedProx across multiple datasets and client configurations, with gains in mIoU, DSC, and convergence speed.
  • The framework requires explicit task-specific design that jointly accounts for geometric consistency and distributional robustness.
  • Training stops after a fixed communication-round budget rather than an explicit convergence criterion.The text proposes monitoring validation Dice or loss plateaus as a future stopping strategy.
  • The experiments assume all clients participate reliably in every communication round, unlike deployments with intermittent institutional availability.Future work is suggested on partial and asynchronous participation, including random client sampling.
  • FedSSL and FedPGD improve optimisation stability and structural fidelity but do not directly address the scarcity of annotated training data.This limitation motivates the subsequent structure-aware synthetic-data generation approach.

Structure-Aware Diffusion for Data-Efficient Catheter and Guidewire Segmentation

The structure-aware diffusion framework generates synthetic fluoroscopic videos that preserve thin catheter and guidewire geometry while improving temporal coherence and perceptual fidelity. These sequences also improve downstream segmentation under limited, heterogeneous, and federated data conditions.

  • Motivation: Existing generative methods struggle to reproduce thin, low-contrast catheter and guidewire structures reliably in X-ray fluoroscopy.Their failures include blurred, missing, fragmented, or temporally inconsistent instrument appearances.
  • Method: The method combines SSL-based structural supervision, domain-adaptive latent modelling, and perceptual objectives within a diffusion framework.It is evaluated on real-animal and phantom fluoroscopic datasets.
  • Downstream Segmentation: Structure-aware synthetic data improve segmentation under limited annotations, data scarcity, and domain shift in both centralised and federated settings.Increasing the synthetic-data ratio from 0.5× to 2.0× and 4.0× produces substantial Dice gains, whereas 0.5× provides only marginal improvement.
  • Method: The proposed framework extends shape-sensitive supervision to video generation, enforcing structural consistency within frames and across time.Its novelty lies in adapting SDM-based supervision to temporally coherent fluoroscopic synthesis.
  • Results: 325.44 to 201.11 FVD on real-animal data and 160.34 to 144.33 on phantom data are achieved relative to Endora.The method maintains competitive FID and IS scores while reducing FVD.
  • Results: 0.8578 to 0.9134 Skeleton Dice on phantom data is achieved over Endora, with improvements in SSIM and Skeleton Dice across both datasets.The gains reflect stronger preservation of thin-instrument geometric continuity.
  • Ablation Study: 191.96 FVD is reached after adding SDM supervision, identifying shape-sensitive loss as the primary driver of temporal-coherence improvement in the ablation.
  • Ablation Study: 47.28 FID and 0.9195 SSIM are obtained in the complete model after LPIPS restores perceptual quality following VAE adaptation.The results indicate that VAE adaptation and LPIPS are complementary, with all components needed for the best overall balance.

6.5 Conclusion

The thesis concludes with a structure-aware diffusion framework for synthetic fluoroscopic video that preserves catheter and guidewire geometry while improving downstream learning under distributed, data-scarce conditions.

  • Structure-aware generation: A structure-aware diffusion framework preserves thin catheter and guidewire structures in synthetic X-ray fluoroscopic video.It combines shape supervision, domain-adaptive transfer learning for the grayscale VAE, and perceptual similarity within a two-phase curriculum.
  • Empirical findings: Experiments on real-animal and phantom datasets produce smoother, more consistent sequences and effective supervision for downstream segmentation under limited annotations and domain shift.
  • Conclusion: Synthetic fluoroscopic generation should optimize both realism and structure preservation rather than realism alone.
  • Limitations: Evaluation is limited by pseudo-label dependence, unquantified run-to-run variance, and testing at 128×128 resolution below native acquisition resolution.
  • Clinical relevance: The framework supports privacy-sensitive collaborative learning by generating data that preserve task-relevant catheter and guidewire geometry.

7.1 Summary

The thesis develops a connected set of methods for thin-instrument segmentation under limited annotation, privacy constraints, heterogeneous federated data, and domain shift, supported by the CathAction benchmark and structure-aware learning.

  • CathAction Dataset and Benchmark: CathAction provides a benchmark spanning phantom, animal, and human fluoroscopic domains with annotations for segmentation and related endovascular tasks.It enables systematic evaluation under inter-domain variation and realistic fluoroscopic difficulty.
  • Shape-Sensitive Loss: Shape-Sensitive Loss represents masks as signed distance maps and compares transformer-derived structural features, reducing fragmentation without changing segmentation backbones.Experiments across multiple backbones showed consistent improvements over standard losses.
  • Federated Learning with Shape-Sensitive Loss: FedSSL extends shape-sensitive supervision to federated learning by aligning local predictions, annotations, and global outputs across heterogeneous client data.It improves segmentation performance and structural preservation relative to federated baselines.
  • Federated Learning with Projected Gradient Descent: FedPGD integrates projected-gradient adversarial optimisation into local federated training to improve robustness to domain variation and imbalanced local data.The framework demonstrated improved convergence, generalisation, and segmentation performance over standard federated baselines.
  • Structure-Aware Diffusion for Synthetic X-ray Video Generation: A latent-diffusion generator combines structural supervision, domain-adaptive VAE transfer learning, perceptual constraints, and a two-phase curriculum for thin-instrument video.Generated data improved downstream segmentation in centralised and federated settings under limited annotations.

7.2 Limitations

The thesis’s evidence is bounded by incomplete acquisition diversity, limited experimental variability reporting, idealised federated participation assumptions, pseudo-label dependence, and reduced-resolution evaluation.

  • General limitation: Run-to-run variance is not quantified, especially in low-data federated settings where method margins are smallest.The reported trends are consistent across two datasets and multiple downstream settings, but missing standard deviations limit statistical strength.
  • Benchmark scope: CathAction does not capture the full diversity of scanner types, institutions, and patient populations encountered clinically.Generalisability claims are therefore bounded by represented acquisition conditions.
  • Federated evaluation: FedSSL and FedPGD use fixed communication-round budgets and assume full, reliable client participation in every round.These assumptions do not reflect institutions that may be intermittently unavailable in realistic federated deployment.
  • Synthetic augmentation: Synthetic augmentation depends on pseudo-labels produced by a model trained on real data, limiting benefits when annotated data are insufficient or domain shift makes labels unreliable.
  • Resolution scope: Generative evaluation at 128×128 resolution does not establish behaviour at full clinical fluoroscopic resolution.

7.3 Future Work

Future work targets broader clinical validation, stronger federated optimisation, extensions beyond segmentation, improved temporal and anatomical generation, annotation efficiency, and workflow integration.

  • Clinical-Scale Federated Validation: Larger multi-centre federated studies should test robustness across scanner types, procedural workflows, institutions, and patient cohorts.Such validation would better reflect the scale and complexity of clinical deployment.
  • Advanced Federated Optimisation: Future federated optimisation could address severe heterogeneity through adaptive aggregation, personalised formulations, uncertainty-aware weighting, and partial participation.
  • Beyond Segmentation: Structure-aware learning could extend from segmentation to tracking, tip localisation, collision prediction, action understanding, and 3D reconstruction.
  • Synthetic Data Generation: Synthetic video generation could improve through stronger temporal modelling, explicit motion supervision, anatomy conditioning, controllability, and joint generation-segmentation.These directions could target rare configurations and clinically important edge cases.
  • Federated Synthetic Training: Generative modelling could be integrated directly into federated training through local site-adapted generators contributing synthetic diversity.This may help mitigate imbalance, improve local generalisation, and reduce dependence on scarce annotations.
  • Structural Supervision: Boundary- and topology-aware losses, centerline supervision, graph representations, and continuity constraints could strengthen structural supervision.
  • Annotation-Efficient Learning: Annotation-efficient research should combine weak, semi-supervised, and self-supervised learning with federated learning and synthetic generation.
  • Clinical Workflow Integration: Future clinical translation must address computational efficiency, hospital infrastructure constraints, clinician trust, interpretability, and privacy-preserving deployment compatibility.

Appendix

The appendix specifies reproducible training configurations for segmentation, federated learning, and diffusion experiments, and describes data preparation and algorithm procedures. It also documents the integration of adversarial optimisation and synthetic data into federated training.

  • Structure-Aware Diffusion Training: Diffusion training used 128 × 128 sequences of 16 frames, AdamW, mixed precision, batch size 1, and 80k iterations.Checkpoints were saved every 10k steps, and the total objective combined diffusion, structural distance-map, VAE, and LPIPS losses.
  • Segmentation Training: Segmentation models used 512 × 512 inputs, SGD with momentum 0.9, polynomial learning-rate decay, and a Dice-plus-SSL objective.The objective weights were γ=0.5 and δ=0.5, selected by grid search on a held-out validation split.
  • Shape-Sensitive Loss: SDM representations were produced with blurred Canny edges and an L2 distance transform, then compared through frozen ImageNet-initialised Vision Transformer features.The ViT used centre-cropped 384×384 inputs and ImageNet normalisation.
  • Federated Learning: Federated experiments used four Dirichlet-partitioned clients, one local epoch per round, 50 rounds, and FedAvg aggregation with SGD-based local optimisation.The partition used β=0.5 to simulate non-IID heterogeneity.
  • Federated Learning: FedPGD added five-step projected-gradient adversarial perturbations, while FedSSL incorporated the shape-sensitive loss into local optimisation.FedPGD independently back-propagated clean and perturbed losses under an ℓ∞ bound of ε=0.01.
  • Synthetic Data Integration: Synthetic diffusion sequences were pseudo-labelled with a U-Net, mixed with real client data using a curriculum schedule, and reported to improve generalisation under data scarcity and domain shift.The federated datasets used representative subsets of approximately 5,000 real-animal frames and 20,000 phantom frames, with approximately 80/20 train-test splits.

B.4 Structure-Aware Diffusion Training

This section identifies the structure-aware diffusion training algorithm and records the thesis’s publication and reuse context. The listed diffusion work was submitted for review as a 2026 workshop contribution.

  • Structure-Aware Diffusion Training: Structure-Aware Diffusion Training is presented as Algorithm 5, taking a training dataset as input and producing a trained diffusion model.The procedure includes a training loop and computes perceptual LPIPS loss.
  • Copyright and Reuse Permissions Appendix: The thesis reuses material from published Springer and IEEE works with corresponding permissions and copyright acknowledgments.The listed reused works include shape-sensitive loss, FedSSL, FedPGD, and CathAction.
  • Under Review / Submitted Work: The structure-aware video diffusion work is listed as a 2026 DEMI workshop submission under review.The submission concerns synthetic X-ray data generation for federated medical image segmentation.
Loading 2609.06876v1…