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

Fluoroscopic sequence annotation is time-consuming and requires specialist expertise, leaving institutions with small datasets compounded by non-IID heterogeneity. This thesis combines structure-aware federated segmentation with synthetic-sequence generation, and reports consistent advantages over FedAvg and FedProx across datasets and client configurations.

  • Problem

    Fluoroscopic sequence annotation is time-consuming and requires specialist expertise, leaving institutions with small datasets compounded by non-IID heterogeneity.

    Source: §1.1 · p. 21

  • Method

    The thesis develops structure-aware federated segmentation using shape-sensitive supervision and adversarial optimisation, alongside structure-aware synthetic fluoroscopic sequence generation to augment limited local training data.

    Source: §1.1 · p. 21, §5.5 · p. 111, §7.1 · p. 130

  • Results

    Both FedSSL and FedPGD consistently outperform FedAvg and FedProx across datasets, client configurations, and communication rounds, with gains in mIoU, DSC, and convergence speed.

    Source: §5.4.6 · p. 110, §5.5 · p. 111

  • Takeaways & Limitations

    Structure-aware synthetic generation is an effective independent mechanism for improving federated segmentation under data scarcity.

    Source: §7.1 · p. 131

    Broader generalisability remains unestablished because further validation is needed across diverse clinical acquisition conditions and the data do not capture full scanner, institutional, and patient diversity.

    Source: §7.2 · p. 132

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 … 1.2 Thesis Scope and Objectives

The thesis develops complementary structure-aware methods for catheter and guidewire segmentation in endovascular fluoroscopy, targeting structural representation, benchmark data, federated robustness, and annotation scarcity. Its objectives progress from structure-aware supervision and benchmark development to federated optimisation and synthetic data generation.

Declaration / Acronyms

  • The declaration states that the thesis is the author's own work submitted solely for a PhD, while the acronym list defines terms including DINO, EnDora, DSC, IoU, and mIoU.The declaration also says published work and reused illustrations are referenced and credited.Source: Declaration · p. 5, Acronyms · p. 17

Introduction

  • The thesis develops structure-aware catheter and guidewire segmentation methods for distributed clinical settings, addressing structural representation, data availability, heterogeneity, and annotation scarcity.The introduction frames the work around structure-aware learning, robustness-oriented federated optimisation, and data-efficient augmentation.Source: Introduction · p. 19

1.2 Thesis Scope and Objectives

  • The scope is a progressive set of complementary methods intended to improve segmentation accuracy and generalisation across distributed clinical data.The motivating limitations include missing structure-aware benchmarks, inadequate structural losses, institutional data fragmentation and heterogeneity, and scarce annotations.Source: §1.2 · p. 23
  • The structural-representation objective is to use signed distance maps and transformer embeddings to capture continuity in thin, elongated instruments.This addresses the limitation of conventional overlap-based objectives, which can yield fragmented predictions despite satisfactory overlap scores.Source: §1.1 · p. 20, §1.2 · p. 23
  • CathAction is intended as a benchmark for catheter and guidewire segmentation, using real-animal and phantom acquisitions to support evaluation under heterogeneous conditions.The motivation describes CathAction as spanning phantom, preclinical animal, and real-world human procedures, addressing limited and narrow existing datasets.Source: §1.1 · p. 21, §1.2 · p. 24
  • The federated-learning objective adapts structure-aware supervision to improve segmentation consistency under Non-IID data distributions.The motivation identifies heterogeneous institutional data and inconsistent local updates as challenges for stable, structurally consistent federated segmentation.Source: §1.1 · p. 21, §1.2 · p. 24
  • The robustness objective integrates adversarial optimisation into federated learning to improve convergence stability and generalisation across heterogeneous clients.This is presented as a robustness-oriented contribution in addition to structure-aware supervision.Source: §1.1 · p. 22, §1.2 · p. 24
  • The data-scarcity objective is to develop a structure-aware diffusion framework that generates synthetic X-ray video for data-limited federated training.The motivation notes that expert annotation is time-consuming and local datasets are small, and calls for structurally consistent synthetic sequences.Source: §1.1 · pp. 21–22, §1.2 · p. 24

1.3 Thesis Contributions · 1.4 List of Publications · 1.5 Outline

The thesis develops an integrated framework for catheter and guidewire analysis that addresses data scarcity, structural continuity, and privacy-preserving federated optimisation. Its chapters progress from dataset development through structure-aware learning and federated methods to synthetic video generation.

1.3 Thesis Contributions

  • Together, the contributions address data availability, structural learning, and privacy-preserving optimisation for catheter and guidewire analysis in fluoroscopic imaging.The thesis introduces CathAction, SSL, FedSSL, FedPGD, and structure-aware diffusion as parts of this framework.Source: §1.3 · p. 25

1.3.1 CathAction: A Large-Scale Benchmark for Endovascular Intervention Analysis

  • CathAction provides approximately 600,000 annotated frames for action recognition and collision detection, plus 40,000 pixel-level masks for catheter and guidewire segmentation.The dataset combines silicon phantom, preclinical animal, and real-world human procedures.Source: §1.3.1 · p. 26

1.3.2 Shape-Sensitive Loss for Catheter and Guidewire Segmentation

  • The Shape-Sensitive Loss encodes masks as signed distance maps and compares Vision Transformer feature embeddings to encourage structural continuity in thin-instrument segmentation.Quantitative and qualitative evaluations on large-scale X-ray datasets report consistent improvements over existing baselines.Source: §1.3.2 · p. 27

1.3.3 FedSSL: Federated Learning with Shape-Sensitive Loss

  • FedSSL applies shape-sensitive supervision in local, global, or combined modes to improve structural consistency and boundary preservation across heterogeneous sites.Experiments on real and phantom X-ray datasets report improvements over standard federated baselines.Source: §1.3.3 · p. 28

1.3.4 FedPGD: Federated Learning with Projected Gradient Descent

  • FedPGD uses projected-gradient-descent adversarial samples aligned with local and global predictions to improve generalisation and convergence under domain heterogeneity.Validation on real and phantom X-ray datasets reports improved convergence and generalisation over established federated baselines.Source: §1.3.4 · p. 29

1.3.5 Structure-Aware Diffusion for Synthetic Fluoroscopic Video Generation

  • The structure-aware diffusion framework generates temporally coherent synthetic X-ray sequences using structural supervision and domain-adaptive reconstruction, supporting segmentation when annotations are limited.Experiments across centralised and federated settings report improved segmentation with generated data, particularly under federated data scarcity and domain shift.Source: §1.3.5 · p. 30

1.4 List of Publications

  • The List of Publications section introduces a summary of the thesis contributions.Source: §1.4 · p. 31

1.5 Outline

  • The thesis outline moves from background and CathAction to SSL, federated methods, synthetic video generation, and concluding findings and future work.The chapters also place the methods in the context of inter-hospital variability, Non-IID data, and limited annotations.Source: §1.5 · p. 32

2.1 Imaging Background for Endovascular Intervention · 2.2 Segmentation in Endovascular Fluoroscopic Imaging

These sections describe the imaging challenges of endovascular fluoroscopy and their implications for catheter and guidewire segmentation. They also address why evaluation of thin-structure segmentation must consider structural continuity beyond region overlap or local boundary deviations.

2.1.1 Interventional X-ray Fluoroscopy

  • Fluoroscopy uses rapid sequential X-ray projections for continuous guidance, prioritizing temporal continuity over clarity; low-dose noise, anatomical overlap, and viewing angle can make instruments faint or discontinuous.The projection collapses three-dimensional anatomy into two dimensions, and dose constraints reduce signal-to-noise ratio [25, 26, 27].Source: §2.1.1 · p. 34, §2.1.2 · pp. 35–36

2.1.3 Operational Variability and Cross-Site Heterogeneity

  • Institutional differences in anatomy, imaging conditions, and acquisition systems produce heterogeneous data distributions, compounding visual ambiguity and making consistent segmentation across sites difficult.Fluoroscopic datasets are also institutionally held, limited in size, and subject to regulatory constraints such as GDPR and HIPAA.Source: §2.1.1 · p. 35, §2.1.3 · p. 37

2.2 Segmentation in Endovascular Fluoroscopic Imaging

  • Segmentation supports tracking, navigation, and procedural guidance, but sparse, faint, elongated targets are fragile: small prediction errors can break connectivity, making geometric continuity essential [31, 34, 35].Low contrast and projection overlap further obscure these curvilinear structures.Source: §2.2 · p. 38, §2.2.2.1 · p. 41

2.2.1.1 Encoder–Decoder CNN Architectures

  • U-Net-style encoder–decoder CNNs are foundational segmentation models, but local receptive fields can miss long-range dependencies along elongated instruments and produce discontinuous predictions.These architectures balance contextual features with spatial localization, while optimisation and loss design address their geometric-continuity limitations.Source: §2.2.1 · p. 38, §2.2.1.1 · p. 39

2.2.1.2 Transformer-Based and Hybrid Architectures

  • Transformers and hybrids model global context through self-attention, but do not explicitly enforce geometric consistency, so overlap scores can mask incoherent thin-structure predictions [40] [41].TransUNet combines transformer encoders with U-Net-style decoders, while Swin architectures use hierarchical attention windows.Source: §2.2.1.2 · p. 40

2.2.2.2 Limitations of Pixel-Wise Optimisation

  • Pixel-wise optimisation is dominated by background pixels in sparse targets, so small centerline errors may have little effect on loss and predictions can remain fragmented despite high accuracy.Figure 2.5 illustrates that fragmented and connected predictions can have comparable Dice scores.Source: §2.2.2.2 · pp. 41–42, Figure 2.5 · p. 41

2.2.3 Loss Functions for Thin-Structure Segmentation

  • Thin-structure loss formulations range from pixel-wise and region-based objectives to boundary-aware and topology-aware methods, addressing imbalance, boundary accuracy, and connectivity [44, 45, 46, 47, 48, 35].Examples include cross-entropy and focal loss, Dice and Tversky losses, Hausdorff and boundary loss, and clDice.Source: §2.2.3 · p. 42

2.2.4 Evaluation Metrics

  • Overlap metrics such as Dice and IoU quantify region agreement but can overlook breaks in thin structures; reliable assessment also needs boundary-deviation and structural-continuity measures.Hausdorff Distance is a distance-based metric for boundary deviation.Source: §2.2.4.1 · p. 43, §2.2.4 · p. 43, §2.2.4.2 · p. 44

2.2.4.2 Distance-Based Metrics

  • HD captures worst-case boundary errors but is outlier-sensitive, while HD95 offers a more robust estimate of boundary deviation.HD measures maximum boundary discrepancy; a single mispredicted point can dominate it.Source: §2.2.4.2 · p. 44
  • Distance-based metrics are local and can miss fragmented or disconnected catheter and guidewire predictions despite small boundary deviations.They do not capture global structural relationships or fully represent geometric consistency.Source: §2.2.4.2 · p. 44
  • Evaluating thin-structure segmentation therefore requires strategies that explicitly account for connectivity and continuity.These properties are critical for elongated targets such as catheters and guidewires.Source: §2.2.4.2 · p. 44

2.3 Federated Learning for Endovascular Fluoroscopic Imaging

This section contrasts centralised and classical distributed training with federated learning (FL), which keeps medical data at their institutions while addressing the non-i.i.d. data distributions and client drift that challenge collaborative training.

2.3.1 Centralised and Distributed Learning Paradigms

  • Medical image segmentation traditionally uses centralised training with data aggregated in one repository, but privacy regulations, institutional constraints, and data ownership can make this infeasible.Source: §2.3.1 · p. 45

2.3.1.1 Centralised Learning

  • Centralised training pools data from multiple sources, benefiting from greater data diversity and generally stable, efficient convergence, but clinical sites face privacy, policy, and ethical barriers to sharing raw data.Source: §2.3.1.1 · p. 46

2.3.1.2 Distributed Learning

  • Classical distributed training partitions data across workers that synchronise gradients or parameters through shared, trusted infrastructure, but clinical data are non-i.i.d. across institutions.Differences in patient populations, imaging equipment, and acquisition protocols undermine the approximate i.i.d. assumption.Source: §2.3.1.2 · pp. 46–47

2.3.2 Federated Learning Framework

  • FL achieves collaborative training without exchanging raw data, using local client optimisation and periodic server aggregation while accommodating non-i.i.d. data distributions.These features are intended for privacy-sensitive settings where data are distributed and statistically heterogeneous.Source: §2.3.2 · p. 47

2.3.3 Federated Averaging (FedAvg)

  • FedAvg alternates local client optimisation with weighted server aggregation, giving clients with larger datasets proportionally greater influence on the global update.More local steps reduce communication but may amplify client drift; more frequent synchronisation improves update alignment at increased communication cost.Source: §2.3.3 · pp. 48–49

2.3.4 Statistical Heterogeneity and Client Drift

  • Under non-i.i.d. client data, local gradients can conflict and accumulate into client drift, slowing convergence and degrading the aggregated global model [57].The section attributes distribution differences to institution-specific patient populations, imaging equipment, and acquisition protocols.Source: §2.3.1.2 · p. 47, §2.3.4 · p. 50

2.4 Adversarial Training · 2.5 Synthetic Data Generation for Endovascular Fluoroscopy

These sections explain adversarial training as a way to optimize robustness against bounded input perturbations, and diffusion models as a means of generating fluoroscopic training data where privacy and limited institutional datasets constrain data availability. They outline DDPM denoising and latent diffusion’s computational advantage for federated settings.

2.4 Adversarial Training

  • Small, bounded perturbations can produce visually similar X-ray images that significantly alter neural-network predictions [58] [59].Figure 2.8 illustrates adding a perturbation δ to an image x to produce x + δ.Source: §2.4 · p. 51, Figure 2.8 · p. 51
  • Adversarial training uses a min–max objective: an inner step finds perturbations that maximize prediction error, while an outer step updates model parameters for robustness [60].Because exact inner maximization is intractable, projected gradient descent (PGD) is commonly used to construct adversarial examples during training [60].Source: §2.4 · pp. 51–52
  • In federated fluoroscopic segmentation, site appearance shifts and client drift motivate adversarial objectives, but robustness may trade off against segmentation accuracy.The passages suggest such objectives may encourage invariant representations across distributional boundaries.Source: §2.4 · p. 52

2.5 Synthetic Data Generation for Endovascular Fluoroscopy

  • Small institutional datasets and privacy restrictions against centralised pooling motivate synthetic data generation to augment clients or broaden coverage across heterogeneous sites.Diffusion models iteratively denoise random noise into structured samples and have demonstrated stronger high-fidelity image synthesis than GANs and VAEs.Source: §2.5 · p. 53

2.5.1 Denoising Diffusion Probabilistic Models

  • DDPMs progressively add Gaussian noise to a clean sample, then learn a reverse denoising process that generates samples from random noise.Figure 2.9 illustrates the forward corruption and learned reverse process.Source: Figure 2.9 · p. 53, §2.5.1 · p. 54

2.5.2 Latent Diffusion Models

  • Latent Diffusion Models apply diffusion in compressed latent space rather than pixel space to address the computational expense of high-dimensional images.An encoder maps an image to a latent representation, diffusion operates on that representation, and a decoder reconstructs the generated image.Source: §2.5.2 · pp. 54–55
  • Latent-space generation reduces computational cost while maintaining high visual fidelity, and can synthesize site-specific fluoroscopic images to augment underrepresented client datasets and reduce heterogeneity [69].This efficiency is relevant to resource-constrained distributed clinical environments.Source: §2.5.2 · p. 55

CathAction: A Large-Scale Benchmark for Endovascular Intervention Analysis · 3.1 Introduction · 3.2 Related Work

CathAction is a multi-source benchmark for endovascular intervention analysis, pairing large-scale annotations with tasks spanning segmentation, action understanding, and collision detection. The chapter benchmarks representative methods and discusses challenges that existing datasets and methods have not addressed.

CathAction: A Large-Scale Benchmark for Endovascular Intervention Analysis

3.1 Introduction

Statement of Contribution

  • The candidate partially annotated ground-truth masks and designed, implemented, and evaluated segmentation, phantom-to-animal domain adaptation, and human-data segmentation tasks.Source: Statement of Contribution · p. 57

3.2 Related Work

  • Existing datasets and methods are constrained by small or private data and narrow task coverage, while CathAction supports segmentation, collision detection, action recognition, and anticipation.Source: §3.1 · p. 58, §3.2 · p. 60

3.1 Introduction

  • CathAction provides manually labeled ground truth for segmentation, action understanding, and collision detection, and benchmarks anticipation, recognition, segmentation, and collision detection.Source: §3.1 · p. 59

Collision Detection.

  • CathAction adds bounding-box annotations for collision events in phantom and real-world data, enabling collision localization and real-time visual or haptic feedback development.Source: Collision Detection · p. 61

3.3 The CathAction Dataset · 3.4 Tasks and Benchmark Evaluation

Together, these sections introduce CathAction’s annotated phantom, animal, and human fluoroscopy data and its benchmarks across five endovascular task families, including held-out human out-of-domain evaluation. The benchmarks reveal substantial domain gaps, with models performing better in-domain than when transferred from phantom and animal data to human data.

3.3.2 Data Annotation

  • Annotations cover temporal action understanding, tip-collision localisation, and pixel-level segmentation, with catheter and guidewire labeled as separate classes.Annotations are aligned at 24 FPS; action labels comprise five classes, and collision boxes distinguish collision from normal.Source: §3.3.2 · p. 64, §3.3.2.1 · p. 65, §3.3.2.2 · p. 66

3.3.3 Dataset Statistics

  • CathAction contains approximately 600,000 annotated frames, around 40,000 catheter and guidewire masks, and 650 videos, making it one of the largest endovascular-intervention datasets.The frames support action understanding and collision detection; the masks support catheter and guidewire segmentation.Source: §3.3.3 · p. 66
  • CathAction spans phantom, animal, and human data, with a large domain gap that makes it a benchmark for evaluating adaptation from controlled data to real-world scenarios.Human patient sequences were retrospectively extracted under institutional approval and data-protection regulations, with identifiers removed.Source: §3.3.1.3 · p. 63, §3.3.1.4 · p. 63, §3.3.3 · p. 67

3.4 Tasks and Benchmark Evaluation

  • Five tasks—anticipation, recognition, segmentation, collision detection, and domain adaptation—are evaluated using phantom and animal data for training and testing, with human data reserved for out-of-domain evaluation.By default, 80% of phantom and animal data are used for training and 20% for testing.Source: §3.4 · p. 68

3.4.1 Catheterisation Anticipation

  • AFFT outperforms RU-LSTM by 2.83% on Accuracy, 2.58% on Precision, and 2.86% on Recall for catheterisation anticipation.Transformer-based methods perform better than CNN- or LSTM-based models in the reported comparison.Source: Table 3.2 · p. 68, §3.4.1 · p. 69

3.4.2 Catheterisation Recognition

  • 62.5% top-1 accuracy on five classes is achieved by TDN with ResNet101, the best reported catheterisation recognition result.The section describes recognition as remaining challenging because instrument appearance is similar across environments and actions depend on visual characteristics.Source: Table 3.3 · p. 70, §3.4.2 · p. 70

3.4.3 Catheter and Guidewire Segmentation

  • SegViT achieves the best catheter and guidewire segmentation performance, although its improvement over other methods is not large.Transformer-based TransUNet and SegViT achieve higher accuracy than traditional UNet.Source: Table 3.4 · p. 71, §3.4.3 · p. 71

3.4.4 Collision Detection

  • Collision detection is defined as object detection of instrument tips, with bounding boxes labeled collision or normal.The benchmark explores YOWO, YOWO-Plus, STEP, HIT, and tiny-object methods Yolov and EFF.Source: §3.4.4 · p. 71
  • 16.91 is the top mean AP for collision detection; tiny object detectors such as Yolov and EFF achieve higher accuracy than other normal detectors.The results remain relatively low, and EFF has difficulty detecting catheter–blood-vessel collisions.Source: §3.4.4 · p. 72, Table 3.5 · p. 73

3.4.5 Domain Adaptation: Phantom to Animal

  • Phantom-to-animal evaluation trains exclusively on phantom data and tests directly on animal data without adaptation, establishing a baseline for domain shift across the benchmark tasks.Phantom, animal, and human data differ in visual appearance, contrast, and anatomical complexity, and real animal or human tasks are more challenging.Source: §3.4.5 · pp. 72–73, §3.4.5.4 · p. 74

3.4.5.1 Anticipation Adaptation

  • Anticipation methods show a significant accuracy drop when evaluated under phantom-to-animal domain adaptation.The passage compares these results with the setup in Table 3.2 and notes challenges in unforeseen catheterisation situations.Source: Table 3.6 · p. 73, §3.4.5.1 · p. 73

3.4.5.2 Recognition Adaptation

  • Recognition accuracy drops approximately 30% when baselines are trained on phantom data and tested on animal data.The passage compares this adaptation setup with the normal setting in Table 3.3.Source: §3.4.5.2 · p. 73, Table 3.7 · p. 74

3.4.5.3 Collision Detection Adaptation

  • Most object detection methods achieve very low accuracy for collision detection under phantom-to-animal domain adaptation.The passages identify a need for methods that can detect collisions in real time.Source: Table 3.8 · p. 74, §3.4.5.3 · p. 74

3.4.5.4 Segmentation Adaptation

  • A significant accuracy drop occurs across phantom-to-animal benchmarks; SegViT still outperforms other segmentation methods.The analysis quantifies domain shift rather than applying a domain-adaptation algorithm.Source: §3.4.5.4 · p. 74, Table 3.9 · p. 75

3.4.6.1 In-domain Human Segmentation

  • Transformer-based models outperform standard UNet on in-domain human segmentation, while SSL shows strong overall performance across evaluation metrics.Models were trained on 80% of human data and tested on the remaining 20%.Source: Table 3.10 · p. 75, §3.4.6.1 · p. 75

3.4.6.2 Phantom + Animal to Human Adaptation

  • Models trained on phantom and animal data show a clear performance drop on human X-ray images; SwinUNet and SSL are relatively more robust.This setting tests models directly on human data, reflecting scenarios where annotated human data is scarce or unavailable.Source: Table 3.11 · p. 76, §3.4.6.2 · p. 76

3.5 Discussion and Conclusion

The discussion presents CathAction as a shared resource for endovascular research and identifies dataset, evaluation, and generalisation limitations. It also describes shape-sensitive loss as a structural approach to thin-instrument segmentation and outlines directions for future work.

3.5 Discussion and Conclusion

  • CathAction provides annotated ground truth for catheter and guidewire segmentation, action understanding, and collision detection across phantom, animal, and real human procedures.Standardised benchmarks, public code, and metrics aim to promote transparency, reproducibility, and collective exploration of endovascular tasks.Source: §3.5 · p. 77
  • The dataset may omit rare clinical scenarios, and benchmarked methods remain insufficiently accurate, with generalisability and real-world adaptability challenges across evaluated tasks.Systematic quality-control procedures such as inter-annotator agreement checks are not explicitly reported for collision and segmentation annotations; vision-derived metrics may not fully reflect clinical needs.Source: §3.5 · p. 77
  • Future work includes domain-adaptive anticipation, recognition, collision detection, and action understanding using graph neural networks, temporal modelling, and transfer learning.The discussion also proposes jointly training multiple endovascular tasks and using CathAction’s scale to train foundation models for endovascular interventions and related medical tasks.Source: §3.5 · p. 77

Shape-Sensitive Loss for Catheter and Guidewire Segmentation

  • SSL encodes masks as signed distance maps, compares transformer-derived shape embeddings with cosine similarity, and consistently improves over baselines on large-scale X-ray datasets.The reported gains include sensitivity to fine structural variations and robustness to scale and translation differences.Source: Shape-Sensitive Loss for Catheter and… · p. 78

4.1 Introduction · 4.2 Related Work · 4.3 Methodology

The chapter proposes a shape-sensitive loss (SSL) for catheter and guidewire segmentation that compares signed-distance-map representations in a learned feature space. It motivates this approach by the difficulty of preserving thin-instrument continuity in fluoroscopic images and by the limits of conventional segmentation losses.

4.1 Introduction

  • Fluoroscopy’s low contrast, noise, and anatomical clutter make thin catheters and guidewires difficult to distinguish, while vessel-wall contact can cause serious complications,,,.Accurate segmentation is described as a prerequisite for computer-assisted intervention systems supporting safer, more precise procedures.Source: §4.1 · p. 79

4.2 Related Work

  • Conventional pixel-, region-, and boundary-level losses do not explicitly preserve continuity or capture global structural relationships in elongated instruments,,.This limitation motivates the proposed structure-aware formulation for fluoroscopic catheter and guidewire segmentation.Source: §4.2 · p. 82

4.3 Methodology

  • SSL addresses discontinuous segmentations by representing masks as signed distance maps (SDMs) and extracting structural features with a Vision Transformer (ViT).The method is designed to account for boundary proximity and global context rather than relying only on discrete pixel labels.Source: §4.3 · p. 83

4.3.1 Preliminaries: Signed Distance Map

  • SDMs encode each pixel’s signed distance to the nearest object boundary, with negative values inside, zero on, and positive values outside the object.This continuous geometric representation provides information beyond binary labels and can make structural misalignment more visible to the learning process.Source: §4.3.1 · pp. 84–85

4.3.2 Feature Extraction using Vision Transformer

  • The ViT extracts high-level embeddings from predicted and ground-truth SDMs, using self-attention to capture long-range geometric dependencies across the image [40].The resulting feature comparison makes the objective sensitive to global geometric consistency rather than only pixel classification.Source: §4.3.2 · pp. 85–86

4.3.3 Shape-Sensitive Loss Computation

  • The SSL measures cosine similarity between predicted and ground-truth SDM feature embeddings, then combines the shape-sensitive term with Dice loss for training.This composite objective couples geometric sensitivity with a conventional segmentation loss.Source: §4.3.3 · p. 87

4.4 Experiments · 4.5 Discussion and Conclusion · Structure-Aware and Robust Federated Learning for Catheter and Guidewire Segmentation

The chapter evaluates shape-sensitive loss (SSL) for catheter and guidewire segmentation across animal and phantom X-rays, then discusses structure-aware federated methods for heterogeneous client data. The experiments report improvements in segmentation performance, while the discussion notes that broader clinical validation remains necessary.

4.4.1 Experimental Configuration

  • Evaluation used annotated animal and phantom CathAction images with a consistent 70:30 split, testing SSL across five segmentation backbones.Animal images were 512 × 512; phantom images were acquired at 1024 × 1024 and cropped to 512 × 512. The backbones were UNet, U-Net++, U-Net3+, TransU-Net, and SwinU-Net.Source: §4.4.1 · p. 88

4.4.2.1 Results on Real Animal X-Ray Images

  • TransU-Net’s Dice coefficient rose from 54.52% to 57.16% (+2.64 percentage points), and it outperformed other architectures in all metrics on real-animal images.All tested segmentation networks benefited from SSL.Source: §4.4.2.1 · p. 88, Table 4.2 · p. 89

4.4.2.2 Results on Phantom X-ray Images

4.4.3 Ablation Study

  • With γ=δ=0.5, TransU-Net reached 57.16% Dice; Dice+SSL beat the standard Cross-Entropy and Dice loss baseline (54.52% to 57.16%), with small weighting variation.The results support complementary contributions from overlap-based supervision and structural guidance.Source: §4.4.3 · pp. 89–90, Table 4.4 · p. 90, Table 4.5 · p. 91
  • Cosine similarity produced the best segmentation performance among the tested measures for comparing feature embeddings extracted from signed distance maps.Source: §4.4.3 · p. 90, Table 4.6 · p. 91

4.4.4 Qualitative Results

  • Across animal and phantom images, SSL preserved continuous, aligned catheter and guidewire structures under faintness or occlusion, with fewer fragments near distal tips and thin guidewire segments.Source: §4.4.4 · p. 91, Figure 4.5 · p. 92

4.5 Discussion and Conclusion

  • SSL converts masks into signed distance maps and uses pretrained ViT shape representations to prioritize geometric continuity; gains were consistent but moderate, requiring validation across more diverse clinical conditions.Source: §4.5 · p. 93

Structure-Aware and Robust Federated Learning for Catheter and Guidewire Segmentation

  • FedSSL adds shape-sensitive loss for structural consistency under heterogeneous client data, while FedPGD uses adversarial optimisation to stabilise convergence; experiments report improvements over FedAvg and FedProx.Federated learning addresses privacy constraints that prevent centralised data sharing, while Non-IID site distributions degrade performance.Source: Structure-Aware and Robust Federated… · p. 94

5.1 Introduction · 5.2 Related Work

This section addresses federated catheter and guidewire segmentation, focusing on preserving thin-structure geometry and robustness across heterogeneous institutions. It introduces complementary shape-sensitive and adversarial approaches and their evaluation against standard federated baselines.

5.1 Introduction

  • The chapter pairs FedSSL’s shape-aware supervision for structural fidelity with FedPGD’s PGD-based adversarial training for robustness under heterogeneous client data.The two methods target complementary needs: geometric consistency and stability under divergent client updates.Source: §5.1 · pp. 95–96

5.2 Related Work

  • Related work leaves thin-structure continuity, federated shape-aware objectives, and PGD-based adversarial training under heterogeneity insufficiently addressed.Standard losses do not adequately penalize fragmentation or discontinuities, while integrating PGD into federated thin-structure segmentation remains largely unexplored.Source: §5.2 · pp. 97–98

5.3.1.1 Motivation

  • FedSSL targets client drift that can produce structurally inconsistent local solutions by adding shape-sensitive supervision alongside ground-truth alignment and guidance from the aggregated model.The motivation is acute for thin, elongated instruments, where small geometric deviations can disrupt segmentation structure.Source: §5.3.1.1 · p. 99

5.3.1.2 Shape-Sensitive Loss

  • The shape-sensitive loss represents masks as signed distance maps to capture boundary and global-shape information, improving tolerance to small misalignments in thin instruments.It is presented as an alternative to standard pixel-wise losses for precise boundary localization.Source: §5.3.1.2 · pp. 99–100

5.3.1.3 Federated Integration of Shape Supervision

  • FedSSL aligns local and global segmentation predictions in shape-aware feature space, with variants for local alignment, global alignment, or both.This couples ground-truth structural accuracy with consistency to the aggregated model, reducing structurally inconsistent updates across clients.Source: §5.3.1.3 · pp. 100–101

5.3.2.1 Motivation

  • FedPGD addresses domain shifts from imaging conditions, anatomy, acquisition protocols, and biased client samples by training with challenging perturbations generated using both local and global models.The approach complements FedSSL, which improves structural fidelity but does not directly address robustness to domain shifts.Source: §5.3.2.1 · p. 101

5.3.2.2 PGD-Based Adversarial Training

  • FedPGD uses two-stage perturbation generation: the local model creates challenging examples, then the global model refines them toward the true-label learning objective.Using both models is intended to make perturbations locally challenging and globally informative for federated optimization under heterogeneous data.Source: §5.3.2.2 · pp. 101–103

5.3.2.3 Federated Training Procedure

  • At each communication round, the server distributes the global model, clients train locally on adversarially perturbed samples, and the server aggregates their updated models by weighted averaging.Adversarial examples are generated on the fly for each mini-batch.Source: §5.3.2.3 · p. 103

5.3.3 Discussion: Complementarity of FedSSL and FedPGD

  • FedSSL adds shape-sensitive supervision to improve structural fidelity, while FedPGD uses client- and global-variability perturbations; PGD maximises model loss rather than only output divergence as VAT does.The FedPGD procedure uses perturbations to refine client model parameters, aiming for stronger adversarial samples and more robust, stable optimisation in heterogeneous federated settings.Source: §5.3.3 · p. 103, §5.3.2.3 · p. 103

5.4.1 Non-IID Federated Training Configuration

  • Experiments used real-animal and phantom X-ray data partitioned across clients with a Dirichlet distribution at β=0.5 to create Non-IID settings.The domains contain 5,086 and 18,791 images, respectively; segmentation used a U-Net backbone, with four- and eight-client configurations evaluated using mIoU and DSC.Source: §5.4.1 · p. 104

5.4.2.1 Evaluation of FedSSL

  • On Real, all three FedSSL variants improve both mIoU and DSC over baselines by round 50, with FedSSL_2 and FedSSL_3 strongest; on combined Real + Phantom, gains appear earlier.Source: §5.4.2.1 · p. 105

5.4.2.2 Evaluation of FedPGD

  • FedPGD improves on FedAvg and FedProx across the Real, Phantom, and combined datasets, achieving the highest Real-dataset mIoU and DSC and the strongest Phantom mIoU.On the combined dataset, it provides the best overall balance; its adversarial optimisation is intended to improve robustness and cross-domain generalisation.Source: §5.4.2.2 · p. 106, Table 5.3 · p. 106, Table 5.4 · p. 107

5.4.3 Scalability with Increasing Client Numbers

  • As clients increase from four to eight, FedSSL_2 and FedSSL_3 maintain strong performance, while FedPGD declines slightly but remains competitive; baselines drop with greater heterogeneity.FedSSL’s global alignment and combined supervision help stabilise learning, whereas FedPGD is strongest in the four-client setting.Source: §5.4.3 · p. 107, Table 5.5 · p. 107

5.4.4 Convergence and Communication Efficiency

  • FedSSL variants improve DSC more rapidly than FedAvg and FedProx, while FedPGD improves mIoU rapidly in early rounds and maintains an advantage over baselines.FedSSL accelerates convergence by aligning structural representations; FedPGD does so through informative adversarial examples. Both show their largest gains early, suggesting plateau-based stopping could reduce communication.Source: §5.4.4 · pp. 108–109, Figure 5.3 · p. 108, Figure 5.4 · p. 108

5.4.5 Qualitative Results

  • FedSSL produces more continuous and accurate delineations around thin boundaries, while FedPGD yields smoother, more coherent masks on low-contrast or ambiguous samples.The qualitative examples are consistent with shape-sensitive supervision improving structural consistency and adversarial optimisation improving robustness to domain variation.Source: §5.4.5 · p. 109, Figure 5.5 · p. 109

5.4.6 Discussion

  • The results support task-specific design for thin-structure segmentation: FedSSL targets structural fragility and FedPGD targets distributional instability, while annotation scarcity remains unresolved.The discussion reports consistent improvements over FedAvg and FedProx across datasets, client configurations, and communication rounds; the remaining data-scarcity constraint motivates the following chapter’s generative approach.Source: §5.4.6 · p. 110

5.5 Conclusion

FedSSL and FedPGD improve federated catheter and guidewire segmentation through complementary structural and adversarial strategies. The conclusion identifies convergence, client participation, and annotated-data scarcity as limitations or directions for further work.

5.5 Conclusion

  • FedSSL and FedPGD consistently outperform FedAvg and FedProx across datasets and client configurations, with gains in mIoU, DSC, and convergence speed.FedSSL preserves geometric structure to improve boundary localisation and mitigate client drift; FedPGD uses PGD-based adversarial optimisation to improve robustness and convergence stability.Source: §5.5 · p. 111
  • Effective federated optimisation for thin-structure segmentation requires task-specific design that addresses both geometric consistency and distributional robustness.FedSSL and FedPGD are presented as complementary approaches addressing these needs.Source: §5.5 · p. 111
  • Training stopped after a fixed communication-round budget, so future work should investigate convergence-based stopping rules such as validation Dice or loss plateaus.These criteria could help determine when training has genuinely converged rather than relying on a preset round count.Source: §5.5 · p. 111
  • The experiments assume reliable participation by every client in every round, unlike deployments with intermittent availability; future work could support partial or asynchronous participation.Suggested approaches include random client sampling, staleness-aware aggregation, and robustness to varying numbers of active participants.Source: §5.5 · p. 111
  • FedSSL and FedPGD do not address annotated-data scarcity, motivating the following chapter’s structure-aware synthetic data generation approach.The conclusion describes annotation scarcity as a fundamental limitation in endovascular imaging.Source: §5.5 · p. 111

Structure-Aware Diffusion for Data-Efficient Catheter and Guidewire Segmentation · 6.1 Introduction

The chapter proposes structure-aware diffusion to generate temporally coherent, structurally faithful synthetic fluoroscopic sequences and uses them to improve catheter and guidewire segmentation under annotation scarcity and domain shift, including in federated learning.

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

  • The chapter frames structure-aware video generation as a response to costly, clinically constrained annotation of thin endovascular instruments in centralized and federated settings.It presents the framework as a strategy for data-efficient endovascular image analysis.Source: Structure-Aware Diffusion for… · p. 112

6.1 Introduction

  • Limited annotations, acquisition variability, and medical-data privacy constraints complicate reliable segmentation and federated training across institutions.Federated learning avoids sharing sensitive patient data, but data scarcity and non-IID institutional distributions remain challenges.Source: §6.1 · p. 113
  • Existing diffusion video generation can miss or fragment thin catheter and guidewire structures, undermining synthetic data’s utility for downstream segmentation.Even small discontinuities can affect instrument localisation and downstream model performance; Fig. 6.1 illustrates the failure with EnDora-generated frames.Source: §6.1 · p. 113, Figure 6.1 · p. 114
  • On real-animal and phantom fluoroscopy data, the framework improved temporal coherence and structural fidelity over a strong diffusion baseline.The evaluation names Fréchet Video Distance (FVD), Structural Similarity Index Measure (SSIM), and Fréchet Inception Distance (FID).Source: §6.1 · pp. 113–114
  • The framework combines SSL-based signed distance map supervision, domain-adaptive latent modelling, and perceptual objectives to generate temporally coherent, structurally consistent sequences.It is designed to preserve thin catheter and guidewire structures.Source: Structure-Aware Diffusion for… · p. 112, §6.1 · p. 113
  • Structure-aware synthetic sequences consistently improved catheter and guidewire segmentation in centralized training and federated learning under data scarcity and domain shift.The chapter presents this as a response to the annotation bottleneck in both centralized and federated settings.Source: Structure-Aware Diffusion for… · p. 112, §6.1 · p. 114

6.2 Related Work

This section covers structure-aware diffusion for fluoroscopic video and its use in synthetic-sequence segmentation supervision, including adaptations for temporal and structural fidelity and evaluations of augmentation in centralised and federated training.

6.2 Related Work

  • Prior work explored diffusion synthesis for medical imaging, [69], and instrument augmentation [211],, but temporally coherent fluoroscopic video remains challenging for thin, low-contrast instruments.Endora established a medical video synthesis baseline, but existing approaches can produce missing or fragmented instruments that limit segmentation utility.Source: §6.2 · p. 115

6.3.1 Baseline Video Diffusion Model; 6.4.1 Video Generation Benchmarks and Evaluation Protocol

  • The method builds on Endora’s latent diffusion backbone, using a VAE, temporal transformer, and U-Net denoiser; evaluation compares it with Endora, StyleGAN-V, and LVDM on real-animal and phantom sequences.The benchmarks assess video coherence and realism alongside structural preservation, with a separate multi-site dataset used for federated downstream validation.Source: §6.3.1 · p. 116, §6.4.1 · p. 120

6.3.2 SDM Loss for Structural Preservation

  • The SDM loss adapts prior segmentation supervision into generative supervision using real-frame structural targets and an inter-frame constraint for temporal coherence.Its novelty is adapting shape-sensitive supervision to temporally coherent fluoroscopic video generation, rather than introducing the SDM–cosine loss itself.Source: §6.3.2 · pp. 116–117

6.3.3 Domain-Adaptive Transfer Learning for X-ray VAE Reconstruction

  • Fine-tuning the VAE on fluoroscopic sequences aligns its latent space with X-ray statistics, improving contrast preservation and reducing reconstruction artifacts; LPIPS complements SDM by restoring texture realism.Without VAE adaptation, diffusion training was unstable and synthesis quality degraded; LPIPS captures texture and intensity discrepancies missed by pixel-level losses.Source: §6.3.3 · p. 118, §6.3.4 · pp. 118–119

6.3.5 Two-Phase Training Curriculum

  • Training shifts from MedSAM structure guidance at iterations 0–60k to DINO refinement at 60k–80k, jointly optimising SDM, VAE, and LPIPS losses; switching earlier harmed structural metrics, while switching later reduced perceptual quality.The objective prioritises geometric consistency with λ_s = 1.0, λ_v = 1.0, and λ_p = 0.4, supporting synthetic-data supervision under annotation scarcity and distributional heterogeneity.Source: §6.3.5 · p. 119, §6.3.6 · p. 119

6.4.2 Quantitative Evaluation of Video Generation Quality

  • FVD fell from 325.44 to 201.11 on real-animal data and from 160.34 to 144.33 on phantom data versus Endora, the lowest scores across both domains.FID and IS remained competitive, indicating improved temporal coherence without sacrificing visual diversity or realism.Source: Table 6.1 · p. 121, §6.4.2 · p. 121
  • On phantom data, Skeleton Dice rose from 0.8578 with Endora to 0.9134, and the proposed method improved SSIM and Skeleton Dice over all baselines on both datasets.The gains are attributed to SDM-based supervision penalising structural discontinuities in thin instruments.Source: Table 6.2 · p. 121, §6.4.2 · p. 121

6.4.3 Ablation Study

  • FVD fell from 325.44 to 287.24 with the two-phase curriculum and to 191.96 after adding SDM supervision, which was the primary driver of temporal-coherence gains.Adding VAE adaptation without LPIPS yielded FID 53.33 and SSIM 0.8835; the complete model achieved best FID 47.28 and SSIM 0.9195.Source: §6.4.3 · p. 122
  • In the ablation study, all four components together achieve the best balance of temporal coherence, structural fidelity, and perceptual quality; VAE adaptation and LPIPS are complementary.VAE adaptation stabilises latent representations under X-ray imaging statistics, while LPIPS restores texture realism and perceptual consistency.Source: §6.4.3 · p. 122, Table 6.3 · p. 122

6.4.4 Downstream Segmentation in Centralised Training

  • With 10% of real data and 10k synthetic videos, both Endora and the proposed augmentation improve Dice and IoU over real-only training, with stronger gains from the proposed method.Segmentation is used as a proxy for whether generated videos provide useful supervision; the generator does not produce masks directly.Source: §6.4.4 · p. 122, Table 6.4 · p. 123

6.4.5 Federated Segmentation Setup

  • The federated evaluation uses offline-generated, pseudo-labelled synthetic data mixed with real samples under a local curriculum, and tests data scarcity and leave-one-site-out generalisation.It uses FedAvg with a U-Net, reports Dice, HD95, and mIoU averaged over runs, and applies late-round Focal-Tversky regularisation for thin-structure boundaries.Source: §6.4.5 · p. 123

6.4.6.1 Data Scarcity Scaling

  • Dice rises from 0.4414 to 0.5088 at 2.0× and 0.5095 at 4.0× synthetic-to-real ratios, with diminishing returns making 2.0× a practical trade-off.A 0.5× ratio provides only marginal improvement over the real-only baseline.Source: §6.4.6.1 · pp. 123–124, Table 6.5 · p. 124

6.4.6.2 Cross-Domain Generalization (LOSO)

  • Across held-out sites, the proposed method achieves the highest Dice and lowest HD95 versus real-only FedAvg and Endora augmentation, but all methods approach zero Dice on severely degraded Site C.Site C yields Dice ≈0.02–0.03 across methods; the proposed method retains a relative improvement, while gains are consistent across Sites A, B, and D.Source: §6.4.6.2 · p. 124, Table 6.6 · p. 125

6.4.7 Qualitative Results

  • Generated sequences show sharper, more temporally stable catheter and guidewire structures than Endora, which often produces blurred or discontinuous appearances.The visual comparison also reports clearer boundary delineation across frames, consistent with the stated Skeleton Dice improvements.Source: §6.4.7 · pp. 124–125, Figure 6.3 · p. 126
  • In federated data-scarcity and LOSO examples, structure-aware augmentation produces smoother, more continuous delineations, particularly on unseen Sites B and C.The real-only baseline often fragments thin-instrument predictions under scarcity and struggles with domain-specific contrast and noise under LOSO.Source: §6.4.7 · p. 125, Figure 6.4 · p. 127

6.5 Conclusion

The thesis connects structure-aware segmentation, federated optimisation, benchmark data, and synthetic fluoroscopic video generation to address data scarcity, heterogeneity, and privacy-sensitive clinical constraints. Its diffusion framework prioritises preserving catheter and guidewire structure as well as visual realism, while limitations remain in pseudo-label availability, run-to-run variability, and resolution.

6.5 Conclusion

  • The structure-aware diffusion framework improves temporal coherence and structural fidelity over a strong Endora baseline on real-animal and phantom data.It combines SDM-based shape supervision, domain-adaptive transfer learning for the grayscale VAE, and a perceptual objective in a two-phase training curriculum.Source: §6.5 · p. 128
  • Synthetic sequences provide effective downstream supervision under limited annotations, including federated settings with data scarcity and domain shift.By preserving task-relevant catheter and guidewire geometry, the generated data are more useful for downstream learning under distributed, privacy-sensitive constraints.Source: §6.5 · p. 128
  • Evaluation leaves pseudo-label dependence, run-to-run variability, and performance at clinical resolution unresolved.Federated augmentation depends on sufficient real data for the labelling model; per-run variance is unreported, and evaluation at 128×128 leaves native-resolution behaviour for future investigation.Source: §6.5 · p. 128
  • The conclusion frames synthetic fluoroscopic generation as a structure-preservation objective, not solely a realism objective.Future work could explore self-supervised or weakly supervised labelling strategies.Source: §6.5 · p. 128

Conclusions

  • The thesis advances catheter and guidewire segmentation through CathAction, SSL, FedSSL, FedPGD, and a structure-aware diffusion model.Together, these contributions address structural representation, benchmark data availability, optimisation robustness, and data efficiency under privacy-sensitive clinical constraints.Source: Conclusions · p. 129

7.1 Summary … 7.4 Final Reflections

The thesis develops connected methods for privacy-constrained catheter and guidewire analysis, combining benchmark data, structure-aware learning, federated optimisation, and synthetic generation. It also identifies limits on the evidence and outlines research needed for broader clinical validation and integration.

7.1 Summary

  • CathAction provides cross-domain annotated benchmarks, while shape-sensitive loss uses signed distance maps and ViT features to improve thin-instrument segmentation without changing backbones.CathAction spans phantom, animal, and human acquisition domains; its benchmark supports evaluation under inter-domain variation and realistic fluoroscopic difficulty.Source: §7.1 · p. 130
  • FedSSL aligns local predictions, annotations, and global outputs to preserve structure across heterogeneous clients, while FedPGD uses local adversarial perturbations and global refinement for robustness to domain variation and imbalanced data.Both federated methods improved segmentation over standard baselines; FedPGD also showed improved convergence and generalisation.Source: §7.1 · pp. 130–131
  • A structure-aware diffusion framework combines SDM supervision, domain-adaptive transfer, perceptual constraints, and staged training to improve video structure and temporal coherence and support downstream segmentation under limited annotations.Generated data improved segmentation in centralized and federated settings, making synthetic generation an independent route to addressing data scarcity.Source: §7.1 · p. 131

7.2 Limitations

  • Across empirical chapters, limited seed reporting leaves run-to-run variability unquantified, especially in low-data federated settings, although gains recur across datasets and downstream settings.The thesis notes that missing per-run standard deviations limit statistical strength despite consistent observed trends.Source: §7.2 · p. 132
  • Other limitations include bounded acquisition diversity, fixed-round training with assumed full client participation, pseudo-label dependence under domain shift, and generative evaluation below native resolution.These conditions qualify generalisation and deployment claims; the thesis presents them as limitations rather than undermining its central findings.Source: §7.2 · pp. 132–133

7.3 Future Work

  • Future work calls for broader multi-centre validation, federated optimisation for stronger heterogeneity and partial participation, and structure-aware learning for tasks beyond segmentation.Proposed extensions include adaptive or personalised federated methods and applications such as tracking, tip localisation, collision prediction, action understanding, and 3D reconstruction.Source: §7.3 · p. 134
  • Further directions include stronger temporal and anatomical video generation, tighter integration of local generators into federated training, topology-aware supervision, annotation-efficient learning, and clinical workflow integration.The thesis also identifies computational efficiency, training stability, clinician trust, interpretability, and privacy-preserving deployment compatibility as translation concerns.Source: §7.3 · pp. 134–135

7.4 Final Reflections

  • The thesis brings together structure-aware supervision, robust federated optimisation, and data-efficient generation to advance privacy-preserving endovascular imaging, while recognizing that clinical deployment challenges remain.Its stated long-term aims include more reliable, generalisable image-guided interventions and potential improvements to patient safety and procedural precision.Source: §7.4 · p. 136

Appendix … B.3 Federated Learning with Projected Gradient Descent (FedPGD)

The appendix documents implementation settings and algorithmic procedures for the diffusion, segmentation, and federated-learning contributions. Its FedPGD procedure combines clean-data and PGD-based adversarial updates during local training.

A.1.2 Training Configuration and Hyperparameters

  • Diffusion training uses 128 × 128 inputs with 16 frames, AdamW at 1 × 10−4, batch size 1, and 80k iterations; the total objective combines diffusion, SDM, VAE, and LPIPS losses.Experiments ran on NVIDIA RTX 8000 or A100 GPUs under Ubuntu 20.04/22.04 using Python 3.9, PyTorch 1.13, and CUDA 11.7.Source: §A.1.2 · p. 156, Table 1 · p. 156, §A.1.1 · p. 156

A.2 Segmentation Model Implementation

  • The segmentation implementation computes signed-distance-map representations using blurred-mask Canny edges and an L2 distance transform, then extracts structural features with a frozen ImageNet-initialised ViT.The ViT input is centre-cropped to 384×384 and normalised with ImageNet statistics.Source: §A.2 · p. 157

A.3 Federated Learning Implementation

  • Federated experiments use a U-Net with FedAvg across 4 Dirichlet-partitioned clients, each training for 1 local epoch per round over 50 rounds.Local optimisation uses SGD with learning rate 0.1, momentum 0.9, weight decay 10−5, and batch size 8.Source: §A.3 · p. 157
  • FedPGD uses 5 PGD attack steps with step size α=0.005 and ℓ∞ perturbation bound ε=0.01; clean and perturbed losses are back-propagated independently.The adversarial samples are generated at each local step using PGD [60].Source: §A.3 · p. 158
  • Synthetic diffusion sequences are pseudo-labelled with a U-Net pretrained on real data, then mixed with real client data using a curriculum schedule.The appendix states this strategy improves generalisation under data scarcity and domain shift; pseudo-labelling follows.Source: §A.3 · p. 158

A.4 Dataset and Annotation Details

  • The real-animal dataset contains 374 videos and approximately 105k frames, while the phantom dataset contains 286 videos and approximately 274k frames.Experiments use approximately 5,000 real-animal frames and 20,000 phantom frames; federated training and test sets are split about 80/20 with non-overlapping client partitions.Source: §A.4 · p. 158

B.1 Shape-Sensitive Loss (SSL)

  • The appendix provides SSL and FedSSL algorithm procedures, including Vision Transformer feature extraction and cosine-similarity computation for SSL, and local-epoch iteration for FedSSL.The implementation-details appendix identifies these specifications as supporting reproducibility of experiments reported in Chapters 4–6.Source: A Implementation Details · p. 156, §B.1 · p. 159, §B.2 · p. 159

B.3 Federated Learning with Projected Gradient Descent (FedPGD)

  • The FedPGD algorithm specifies a standard clean-data update and a separate adversarial update via PGD within its local training procedure.The algorithm is presented with an initial global model, clients, PGD steps, step size, and perturbation bound as inputs.Source: §B.3 · p. 160

B.4 Structure-Aware Diffusion Training · C Copyright and Reuse Permissions

The appendix gives fragments of a structure-aware diffusion-training algorithm and documents the permissions, acknowledgments, and works associated with reused material. It also lists a submitted diffusion paper.

B.4 Structure-Aware Diffusion Training

  • Algorithm 5 trains a diffusion model on a training dataset through repeated steps and computes perceptual LPIPS loss.The supplied algorithm passages show only these elements of the training procedure.Source: §B.4 · p. 161

C Copyright and Reuse Permissions

  • The thesis states that previously published material is reused under publisher policies, with citation and acknowledgment.It identifies Springer Nature and IEEE as publishers whose reuse policies permit material to be reused in theses and dissertations.Source: C Copyright and Reuse Permissions · p. 162
  • Springer Nature reuse follows its author rights policy, which permits Version of Record material in theses and dissertations.The stated acknowledgment is “Reproduced with permission from Springer Nature.”Source: C Copyright and Reuse Permissions · p. 162
  • IEEE reuse follows its author rights guidelines for theses and dissertations, subject to including appropriate copyright notices.The appendix specifies the notice: “© 2024 IEEE. Reprinted, with permission, from the original publication.”Source: C Copyright and Reuse Permissions · p. 162
  • The reuse list includes works on shape-sensitive segmentation, FedSSL, FedPGD, and the CathAction benchmark.These are listed as published or preprint works, with Springer Nature or IEEE acknowledgments shown for the first three.Source: C Copyright and Reuse Permissions · pp. 162–163
  • The appendix lists a 2026 structure-aware video-diffusion paper on synthetic X-ray data generation for federated medical image segmentation as under review.It is listed for the 4th Workshop on Data Engineering in Medical Imaging, held with MICCAI 2026.Source: C Copyright and Reuse Permissions · p. 163

Related papers

Loading 2609.06876v1…