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

Chayun Kongtongvattana

Figure 1.1 : CathAction dataset, a large-scale dataset for endovascular interventions, comprising data acquired from silicon phantom models, animal trials, and real-world human procedures. The top row shows an example X-ray sequence. The bottom row shows ground-truth labels. The catheter is in blue colour , the guidewire is in dark yellow , and the red box indicates the collision of the tip with the blood vessel, while the green box indicates no collision.

Figure 1.2 : Overview of the Shape-Sensitive Loss based on SDM embedding and transformer representations.

Figure 1.3 : FL with SSL framework.

Figure 1.4 : FedPGD framework integrating adversarial optimisation into FL .

Figure 1.5 : Structure-aware synthetic X-ray video generation

Figure 2.1 : Representative examples of medical imaging modalities used in clinical practice, including X-ray, CT, MRI, PET, and ultrasound. Images reproduced from Radiopaedia.org (X-ray: Usman Bashir, rID: 18394; CT: Craig Hacking, rID: 62533; MRI: Vikas Shah, rID: 49349; PET: David Little, rID: 77382; US: Henry Knipe, rID: 38664), under Creative Commons licenses.

Figure 2.2 : Representative fluoroscopic frame illustrating challenges for automated segmentation. Thin instruments (catheter/guidewire), low-contrast regions, and overlapping anatomical structures complicate reliable delineation under projection imaging and dose constraints.

Figure 2.3 : Representative interventional X-ray fluoroscopic frames acquired under different institutional or procedural settings. Variations in contrast distribution, anatomical visibility, and instrument clarity illustrate systematic appearance shifts across acquisition environments, reflecting non-identically distributed ( Non-IID ) data characteristics.

Figure 2.4 : Generic encoder–decoder segmentation architecture following the U-Net design proposed by [ 36 ] . The encoder progressively extracts hierarchical feature representations, while the decoder restores spatial resolution via upsampling and skip connections.

Figure 2.5 : Comparison of thin-structure segmentation outcomes. Although the fragmented prediction achieves a Dice score comparable to the connected prediction, it contains multiple disconnected components, illustrating that overlap-based metrics alone do not guarantee structural continuity.

Category	Representative Loss	Key Property
Distribution-Based	Binary Cross-Entropy, Focal Loss	Treat segmentation as independent pixel-wise classification without explicit spatial modelling.
Region-Based	Dice, Tversky	Optimise region-level overlap between predicted and ground-truth masks.
Boundary-Based	Hausdorff Distance Loss, Boundary Loss	Incorporate boundary distance and contour alignment information.
Topology-Aware	clDice	Encourage connectivity preservation in tubular structures.
Feature-Level / Shape-Sensitive	Shape-Sensitive Loss ( SSL )	Compare structural representations in feature space rather than pixel masks.

Table 2.1 : Conceptual taxonomy of loss functions used in medical image segmentation.

Figure 2.6 : Comparison between centralised and classical distributed training paradigms. Centralised learning aggregates raw data into a single location, whereas distributed training partitions data across multiple workers but assumes a shared and trusted computational infrastructure.

Figure 2.7 : Federated learning workflow. A central server broadcasts the global model to participating clients. Each client performs local training using its private dataset and returns updated model parameters for aggregation. Raw medical data remain local at each institution.

Figure 2.8 : Illustration of adversarial perturbation. A small bounded perturbation

\delta

is added to the original X-ray image

x

to produce an adversarial example

x+\delta

.

Figure 2.9 : Illustration of diffusion-based generative modelling. In the forward diffusion process

q(x_{t}\mid x_{t-1})

, a clean image

x_{0}

is progressively corrupted by Gaussian noise until it becomes nearly pure noise

x_{T}

. A neural network is trained to learn the reverse denoising process

p_{\theta}(x_{t-1}\mid x_{t})

, which gradually removes noise and generates a sample

\hat{x}_{0}

.

Dataset	Modality	Format	#Frames	Source	Tasks
Barbu et al. [ 12 ]	X-ray	Video	535	Human	Segmentation
Wu et al. [ 95 ]	3D Echo	Video	800	Human	Segmentation
Ambrosini et al. [ 3 ]	X-ray	Image	948	Human	Segmentation
Mastmeyer et al. [ 13 ]	3D MRI	Image	101	Human	Segmentation
Yi et al. [ 92 ]	X-ray	Image	2,540	Synthetic	Segmentation
Nguyen et al. [ 80 ]	X-ray	Image	25,271	Phantom	Segmentation

Table 3.1: Comparison of CathAction with existing endovascular intervention datasets.

Figure 3.1 : Silicon phantom model and experimental data collection setup under X-ray fluoroscopy.

Figure 3.2 : Example X-ray fluoroscopic frames from phantom (top), animal (middle), and human (bottom) domains. Instrument visibility decreases and anatomical complexity increases from phantom to human data.

Figure 3.3 : Examples of annotation types in CathAction: (a) temporal action annotation, (b) collision detection with bounding boxes centred on the distal instrument tip, and (c) pixel-level segmentation annotation of catheter and guidewire. The catheter is in red colour and the guidewire is in green colour.

Figure 3.4 : Distribution of annotated catheterisation action classes in the CathAction dataset.

Baseline	Venue	Accuracy	Precision	Recall
CNN [ 112 ]	CVPR 2018	28.98	30.14	29.76
RNN [ 112 ]	CVPR 2018	29.64	30.38	30.44
RU-LSTM [ 113 ]	CVPR 2019	35.08	34.29	34.77
TempAggRe [ 114 ]	ECCV 2020	34.64	35.56	34.71
Trans-SVNet [ 118 ]	IJCARS 2022	29.06	19.67	20.28
AFFT [ 115 ]	WACV 2023	37.91	36.87	37.63

Table 3.2 : Catheterisation anticipation results on CathAction dataset. Values are reported in percentage (%).

Figure 3.5 : Qualitative catheterisation anticipation results. Green indicates correct prediction and red indicates incorrect prediction.

Baseline	Venue	Accuracy	Precision	Recall
TDN-ResNet50 [ 120 ]	CVPR 2021	58.34	59.12	57.22
TDN-ResNet101 [ 120 ]	CVPR 2021	62.50	61.89	62.77
Video Swin Transformer [ 121 ]	CVPR 2022	51.67	52.14	51.24
BEVT [ 122 ]	CVPR 2022	49.28	50.27	49.92
CycleACR [ 123 ]	TPAMI 2025	60.42	58.15	61.24

Table 3.3 : Catheterisation recognition results on CathAction dataset. Values are reported in percentage (%).

Baseline	Dice Score	Jaccard Index	mIoU	Accuracy
U-Net [ 36 ]	51.69	57.51	31.17	63.26
TransUNet [ 41 ]	56.52	55.93	34.13	55.61
SwinUNet [ 135 ]	61.26	59.54	39.53	76.60
SSL [ 228 ]	56.95	56.87	40.87	72.24
SegViT [ 137 ]	63.47	54.12	42.48	68.73

Table 3.4 : Segmentation results on CathAction dataset. Values are reported in percentage (%).

Figure 3.6 : Qualitative results for the collision detection task. The first two columns visualise the collision results, the third column visualises no collision cases, and the last column visualises a failure case where the tip was not detected.

	AP			mAP
Baseline	Collision	Normal	Mean	Collision	Normal	Mean
STEP [ 141 ]	7.79	11.21	10.98	6.92	11.29	9.08
YOWO [ 139 ]	8.32	12.18	11.73	7.46	12.28	9.92
YOWO-Plus [ 140 ]	8.92	12.23	11.77	7.86	12.48	10.28
HIT [ 142 ]	9.37	12.74	12.14	8.18	12.72	10.81
Yolov ∗ [ 143 ]	12.30	21.08	15.89	11.88	20.04	14.11

Table 3.5 : Collision detection results on CathAction dataset. ( ∗ ) denotes tiny object detectors.

Baseline	Venue	Accuracy	Precision	Recall
RU-LSTM [ 113 ]	CVPR 2019	22.93	23.91	22.57
TempAggRe [ 114 ]	ECCV 2020	17.16	18.41	18.23
Trans-SVNet [ 118 ]	IJCARS 2022	19.06	17.67	19.58
AFFT [ 115 ]	WACV 2023	25.67	26.29	26.33

Table 3.6 : Catheterisation anticipation results under domain adaptation (trained on phantom, tested on animal).

Baseline	Venue	Accuracy	Precision	Recall
TDN-ResNet50 [ 120 ]	CVPR 2021	24.19	23.17	24.56
TDN-ResNet101 [ 120 ]	CVPR 2021	25.62	24.52	25.68
Video Swin Transformer [ 121 ]	CVPR 2022	28.79	27.98	28.12
BEVT [ 122 ]	CVPR 2022	31.22	30.48	31.79

Table 3.7 : Catheterisation recognition results under domain adaptation (trained on phantom, tested on animal).

	AP			mAP
Baseline	Collision	Normal	Mean	Collision	Normal	Mean
STEP [ 141 ]	1.53	2.12	1.87	1.09	1.98	1.62
YOWO [ 139 ]	2.12	4.11	3.09	1.97	3.68	2.92
YOWO-Plus [ 140 ]	1.18	1.43	1.21	1.07	1.26	1.09
HIT [ 142 ]	1.31	1.19	1.24	1.06	1.18	1.11
Yolov ∗ [ 143 ]	7.31	8.92	8.09	6.28	7.49	7.21

Table 3.8 : Collision detection results under domain adaptation setup. All methods are trained on phantom data and tested on animal data. ( ∗ ) denotes tiny object detectors.

Baseline	Dice Score	Jaccard Index	mIoU	Accuracy
UNet [ 36 ]	26.58	31.38	12.13	46.07
TransUNet [ 41 ]	16.16	24.19	17.23	33.61
SwinUNet [ 135 ]	17.41	38.14	7.52	40.79
SSL [ 228 ]	26.91	32.04	18.72	42.44
SegViT [ 137 ]	30.74	32.22	11.46	50.00

Table 3.9 : Domain adaptation segmentation results. All methods are trained on phantom and tested on animal data.

Baseline	Dice Score	Jaccard Index	mIoU	Accuracy
UNet [ 36 ]	51.93	35.07	67.31	49.21
TransUNet [ 41 ]	58.39	43.64	71.59	63.66
SwinUNet [ 135 ]	56.60	41.43	70.48	63.23
SSL [ 228 ]	58.88	45.02	72.01	71.48
SegViT [ 137 ]	48.72	34.64	66.99	56.81

Table 3.10 : Human X-ray segmentation results under in-domain setting.

Baseline	Dice Score	Jaccard Index	mIoU	Accuracy
UNet [ 36 ]	24.33	13.85	56.51	26.88
TransUNet [ 41 ]	24.42	15.26	57.28	30.14
SwinUNet [ 135 ]	25.81	15.92	57.58	34.99
SSL [ 228 ]	24.97	16.04	58.75	34.10
SegViT [ 137 ]	26.05	14.97	57.09	26.91

Table 3.11 : Domain adaptation segmentation results on human X-ray images. Phantom and animal data are used for training. All human data are used for testing.

Figure 4.1 : Illustration of the challenge in catheter and guidewire segmentation from fluoroscopic images. Thin elongated structures, low contrast, and background clutter can lead to discontinuous segmentation results, motivating the need for a structure-aware loss function.

Distribution-Based Loss
Binary Cross-Entropy [ 161 ]	$-\sum y\log(p)-(1-y)\log(1-p)$
Weighted Cross-Entropy [ 162 ]	$-\sum w_{y}y\log(p)$
Balanced Cross-Entropy [ 163 ]	$-\beta\sum y\log(p)-(1-\beta)\sum(1-y)\log(1-p)$
Focal Loss [ 44 ]	$-\sum(1-p)^{\gamma}y\log(p)$
Region-Based Loss
Dice Loss [ 50 ]	$1-\frac{2\sum(y\cap p)}{\sum y+\sum p}$

Table 4.1 : Representative loss functions used in semantic segmentation.

Figure 4.2 : Illustration of the generation process of Signed Distance Maps ( SDM ). Top row: Ground-truth segmentation masks. Middle row: Contours extracted from the segmentation masks. Bottom row: Corresponding Signed Distance Maps computed based on the extracted contours.

Figure 4.3 : Overview of SSL computation and its integration within the training pipeline. Predicted and ground-truth segmentation masks are converted into Signed Distance Maps (SDMs), which are then processed by a feature encoder to extract high-level representations. Cosine similarity between the predicted and ground-truth SDM features is computed to measure structural alignment, and the resulting SSL is used to guide model optimisation during training.

Network	Dice	Jaccard	mIoU	Accuracy
TransU-net	54.52	43.87	61.15	78.25
TransU-Net w/ours	57.16	46.04	62.66	79.57
U-Net	46.20	36.62	54.14	71.57
U-Net w/ours	47.95	38.79	55.05	72.89
U-Net++	48.77	39.22	56.17	72.30
U-Net++ w/ours	50.35	41.39	57.08	73.62

Table 4.2 : Comparing results for real animal X-ray images

Network	Dice	Jaccard	mIoU	Accuracy
TransU-net	40.83	33.94	46.01	66.18
TransU-Net w/ours	43.71	36.65	47.21	67.58
U-Net	34.51	27.62	43.14	60.57
U-Net w/ours	37.13	28.56	45.83	62.29
U-Net++	35.24	27.43	44.72	61.76
U-Net++ w/ours	37.25	28.94	45.97	63.12

Table 4.3 : Comparing results for phantom X-ray images

Figure 39

Table 40

Figure 4.5 : Catheter and guidewire segmentation in X-ray images. Row 1: Raw animal X-ray images. Row 2: Segmentation results for animal data. Row 3: Raw phantom images. Row 4: Segmentation results for phantom data. Red indicates the catheter and green indicates the guidewire.

Figure 5.1 : Overview of the FedSSL framework. Local models are guided by both ground-truth masks and global model predictions using an SSL .

Figure 5.2 : Overview of the FedPGD framework. Adversarial examples are generated using a two-stage process involving both local and global models.

Method	Real				Phantom
	25 Rounds		50 Rounds		25 Rounds		50 Rounds
	mIoU	Dice	mIoU	Dice	mIoU	Dice	mIoU	Dice
FedAVG [ 16 ]	32.92	49.18	45.80	61.63	17.61	49.10	18.81	50.24
FedProx [ 56 ]	28.17	35.36	39.70	48.73	15.17	46.29	17.19	48.68
FedSSL_1 ( Ours )	28.32	48.25	48.82	63.48	17.13	48.66	21.03	51.36
FedSSL_2 ( Ours )	28.50	47.59	48.95	63.78	17.08	48.62	19.73	50.85

Table 5.1 : Performance of FedSSL variants compared with baseline FL methods on Real and Phantom datasets.

Method	Real + Phantom
	25 Rounds		50 Rounds
	mIoU	Dice	mIoU	Dice
FedAVG [ 16 ]	41.28	56.19	41.40	58.65
FedProx [ 56 ]	33.53	44.34	37.04	49.65
FedSSL_1 ( Ours )	41.21	56.79	42.31	59.42
FedSSL_2 ( Ours )	41.40	57.14	42.61	59.97

Table 5.2 : Performance of FedSSL variants compared with baseline FL methods on the combined Real + Phantom dataset.

Method	Real				Phantom
	25 Rounds		50 Rounds		25 Rounds		50 Rounds
	mIoU	Dice	mIoU	Dice	mIoU	Dice	mIoU	Dice
FedAVG [ 16 ]	32.92	49.18	45.80	61.63	17.61	49.10	18.81	50.24
FedProx [ 56 ]	28.17	35.36	39.70	48.73	15.17	46.29	17.19	48.68
FedPGD ( Ours )	43.86	58.05	47.67	62.90	17.71	48.40	21.12	50.82

Table 5.3 : Performance of FedPGD variants compared with baseline FL methods on Real and Phantom datasets.

Method	Real + Phantom
	25 Rounds		50 Rounds
	mIoU	Dice	mIoU	Dice
FedAVG [ 16 ]	41.28	56.19	41.40	58.65
FedProx [ 56 ]	33.53	44.34	37.04	49.65
FedPGD ( Ours )	41.38	55.78	42.43	58.95

Table 5.4 : Performance of FedPGD variants compared with baseline FL methods on the combined Real + Phantom dataset.

Method	#Client = 4				#Client = 8
Round	25	25	50	50	25	25	50	50
Metric	mIoU	Dice	mIoU	Dice	mIoU	Dice	mIoU	Dice
FedAVG [ 16 ]	32.92	49.18	45.80	61.63	30.32	46.81	44.18	60.21
FedProx [ 56 ]	28.17	35.36	39.70	48.73	28.48	45.38	42.61	54.25
FedSSL_1 ( Ours )	34.56	51.12	48.82	63.48	31.56	48.21	47.24	61.89
FedSSL_2 ( Ours )	35.23	51.64	48.95	63.78	31.79	48.67	47.43	62.26

Table 5.5 : Performance comparison of different methods under different numbers of clients.

Figure 5.3 : DSC across communication rounds for FedSSL and baseline methods.

Figure 5.4 : mIoU across communication rounds for FedPGD and baseline methods.

Figure 5.5 : Qualitative comparison of segmentation results. Columns show the original X-ray fluoroscopic image, ground truth, FedAvg , FedProx , FedSSL_1, FedSSL_2, FedSSL_3, and FedPGD , respectively.

Figure 6.1 : Comparison between real and EnDora -generated X-ray fluoroscopic frames. While catheter and guidewire structures are visible in the real image, EnDora may fail to reproduce these thin structures, leading to missing or indistinguishable instruments.

Figure 6.2 : Overview of the proposed structure-aware diffusion framework for synthetic X-ray video generation. Building on Endora’s latent diffusion backbone, three domain-specific objectives are incorporated: a SSL that preserves thin catheter and guidewire structures, a domain-adaptive transfer learning objective for the Variational Autoencoder (VAE) that stabilises reconstructions under X-ray imaging statistics, and a perceptual Learned Perceptual Image Patch Similarity (LPIPS) loss that restores realistic texture details. Training follows a two-phase curriculum (MedSAM warm-up

\rightarrow

Self-Distillation with No Labels (DINO) refinement), which first captures structural priors and then consolidates perceptual realism. These design choices produce temporally coherent and structurally consistent X-ray sequences, as validated quantitatively in Sec. 6.4 .

Method	Real			Phantom
Method	FVD	FID	IS	FVD	FID	IS
LVDM [ 224 ]	307.01	79.13	1.18	158.73	138.16	1.53
StyleGAN-V [ 223 ]	432.87	153.12	1.03	255.89	235.64	1.44
Endora [ 216 ]	325.44	46.89	1.21	160.34	86.55	1.60
Ours	201.11	47.28	1.21	144.33	68.09	1.64

Table 6.1 : Fidelity metrics on real-animal and phantom X-ray datasets. We report FVD for temporal coherence and FID/IS for visual realism and diversity. Our method substantially lowers FVD while maintaining competitive FID and IS.

Method	Real			Phantom
Method	SSIM	W-SSIM	SD	SSIM	W-SSIM	SD
LVDM [ 224 ]	.8917	.8911	.8467	.9488	.9475	.8304
StyleGAN-V [ 223 ]	.8528	.8468	.8114	.9075	.9008	.8036
Endora [ 216 ]	.9170	.9118	.8722	.9667	.9646	.8578
Ours	.9195	.9142	.8738	.9863	.9849	.9134

Table 6.2 : Structure preservation metrics on real-animal and phantom X-ray datasets. SSIM, W-SSIM, and Skeleton Dice (SD) emphasise continuity of thin catheter structures. Our method improves structural fidelity, particularly on the phantom dataset.

2PC	SDM	VAE	LPIPS	FVD $\downarrow$	FID $\downarrow$	SSIM $\uparrow$
✗	✗	✗	✗	325.44	46.89	.9170
✓	✗	✗	✗	287.24	53.88	.9191
✓	✓	✗	✗	191.96	99.34	.9137
✓	✓	✓	✗	213.04	53.33	.8835
✓	✓	✓	✓	201.11	47.28	.9195

Table 6.3 : Ablation study on the real-animal dataset. The first row corresponds to the original Endora baseline. 2PC denotes the two-phase curriculum (MedSAM warm-up

\rightarrow

DINO refinement). SDM refers to the structure-aware distance map loss. The full model achieves the best balance between temporal fidelity, perceptual quality, and structural similarity.

Training Setup	Dice	IoU(FG)	mIoU	HD95
Real-animal dataset
Real (10%)	.3661	.2529	.6234	23.4963
Real (10%) + Endora [ 216 ]	.4734	.3339	.6641	21.9784
Real (10%) + Ours	.5188	.3720	.6835	20.6756
Real (100%)	.6929	.5460	.7711	3.0879
Phantom dataset

Table 6.4 : Centralised downstream segmentation with synthetic augmentation. Training with 10% of real data benefits from adding 10k synthetic videos. Both Endora and our method improve over the baseline, with ours achieving stronger gains.

Synthetic ratio	Dice $\uparrow$	HD95 (px) $\downarrow$	mIoU $\uparrow$
Real-only	.4414	18.26	.3280
0.5 $\times$	.4407	17.38	.3309
1.0 $\times$	.4827	21.31	.3482
2.0 $\times$	.5088	19.95	.3615
4.0 $\times$	.5095	21.52	.3529

Table 6.5 : Federated performance under varying synthetic-to-real ratios per client. Best results in bold.

Table 59

Figure 6.3 : Visual comparison of synthetic X-ray generation quality across different methods. Each row corresponds to a different input sequence, while columns show (from left to right): raw input frames, LVDM, StyleGAN-V, Endora, and the proposed method. White dashed boxes indicate regions of interest containing catheter and guidewire structures. Selected examples are further highlighted, where green boxes denote clear and continuous structures, and red boxes indicate missing or degraded structures.

Figure 6.4 : Qualitative comparison under data-scarcity and cross-domain (LOSO) scenarios. The first two rows show results from the data-scarcity experiment. The third and fourth rows correspond to the LOSO setting, evaluated on unseen domains (Site B and Site C). Colour legend: green – ground truth, blue – Real-only, red – Ours.

Parameter	Symbol	Value
Learning rate	–	$1\times 10^{-4}$
Gradient accumulation steps	–	3
Training iterations	–	80k
Input resolution	–	$128\times 128$
Sequence length	–	16 frames
SDM loss weight	$\lambda_{s}$	1.0

Parameter	Symbol	Value
Number of clients	$K$	4
Communication rounds	$T$	50
Local epochs per round	–	1
Local batch size	–	8
Local learning rate	–	0.1
Weight decay	–	$10^{-5}$

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

Abstract

Figures & tables

Appendix

Medical Image Segmentation

Federated Learning for Cross-Modality Medical Image Segmentation via Augmentation-Driven Generalization

SplitFed-CL: A Split Federated Co-Learning Framework for Medical Image Segmentation with Inaccurate Labels

Projection Guided Personalized Federated Learning for Low Dose CT Denoising

Abstract

Medical Image Segmentation

Explore similar work

Federated Learning for Cross-Modality Medical Image Segmentation via Augmentation-Driven Generalization

SplitFed-CL: A Split Federated Co-Learning Framework for Medical Image Segmentation with Inaccurate Labels

Projection Guided Personalized Federated Learning for Low Dose CT Denoising