Projection Guided Personalized Federated Learning for Low Dose CT Denoising

Anas Zafar; Muhammad Waqas; Amgad Muneer; Rukhmini Bandyopadhyay; Jia Wu

eess.IVMar 12, 2026

Projection Guided Personalized Federated Learning for Low Dose CT Denoising

Authors: Anas Zafar, Muhammad Waqas, Amgad Muneer, Rukhmini Bandyopadhyay, Jia Wu

Organizations: The University of Texas MD Anderson Cancer Center, Houston, TX, USA

Abstract

Low-dose CT (LDCT) reduces radiation exposure but introduces protocol-dependent noise and artifacts that vary across institutions. While federated learning enables collaborative training without centralizing patient data, existing methods personalize in image space, making it difficult to separate scanner noise from patient anatomy. We propose ProFed (Projection Guided Personalized Federated Learning), a framework that complements the image space approach by performing dual-level personalization in the projection space, where noise originates during CT measurements before reconstruction combines protocol and anatomy effects. ProFed introduces: (i) anatomy-aware and protocol-aware networks that personalize CT reconstruction to patient and scanner-specific features, (ii) multi-constraint projection losses that enforce consistency with CT measurements, and (iii) uncertainty-guided selective aggregation that weights clients by prediction confidence. Extensive experiments on the Mayo Clinic 2016 dataset demonstrate that ProFed achieves 42.56 dB PSNR with CNN backbones and 44.83 dB with Transformers, outperforming 11 federated learning baselines, including the physics-informed SCAN-PhysFed by +1.42 dB.

Figures & tables

Figure 1: ProFed framework: Architectural overview for projection-guided personalized federated learning.

	Client #1		Client #2		Client #3		Client #4		Client #5		Client #6		Client #7		Client #8		Average
	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM
FedAvg [ 31 ]	29.99	82.20	32.82	78.23	34.68	86.99	32.30	80.63	36.87	92.06	37.36	89.02	32.73	80.66	35.46	87.65	34.03	84.70
FedProx [ 23 ]	29.94	83.12	32.69	75.46	36.95	91.58	32.32	80.91	38.95	95.34	37.58	88.13	32.94	81.62	35.86	88.65	34.65	85.60
FedNova [ 37 ]	29.95	82.29	32.57	80.57	37.77	92.16	32.09	78.98	34.42	88.69	36.05	89.57	32.81	81.41	35.89	88.71	33.94	85.30
MOON [ 22 ]	32.50	75.61	35.55	85.41	35.72	80.30	35.69	82.10	35.38	78.47	36.31	84.35	35.70	81.17	35.62	80.40	35.31	80.98
FedDG [ 28 ]	33.37	81.40	34.01	82.39	33.56	76.37	34.43	81.49	34.06	76.70	33.97	80.17	34.75	80.87	34.25	79.23	34.05	79.83

Table 1: CNN-based quantitative performance (PSNR/SSIM)

	Client #1		Client #2		Client #3		Client #4		Client #5		Client #6		Client #7		Client #8		Average
	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM
FedAvg [ 31 ]	34.31	87.88	38.06	90.54	43.51	97.56	41.05	95.43	45.24	98.47	43.15	97.09	41.45	95.95	42.88	96.16	41.21	94.89
FedProx [ 23 ]	34.32	87.01	37.74	88.89	43.73	97.95	41.12	93.92	45.39	98.61	42.99	97.08	41.38	94.31	42.10	94.55	41.10	94.04
FedNova [ 37 ]	29.55	78.44	31.38	68.13	37.07	82.47	31.79	73.45	40.61	88.80	35.46	79.77	32.38	73.66	35.77	82.22	34.25	78.37
MOON [ 22 ]	34.28	93.20	38.09	91.17	43.64	97.77	41.09	95.66	45.25	98.43	42.82	96.91	41.42	95.97	42.20	96.18	41.10	95.66
FedKD [ 41 ]	35.92	89.93	38.70	91.21	44.26	98.13	42.30	95.99	45.93	98.75	43.57	97.41	42.54	96.22	43.51	96.23	42.09	95.48

Table 2: Transformer-based quantitative performance (PSNR/SSIM)

Figure 2: Qualitative results of six selected comparison methods and our method across different clients using the classical convolutional-based LDCT imaging network. Rows one to five represent Clients #2, #3, #5, #6, and #7, respectively

	Unseen Client #1		Unseen Client #2		Unseen Client #3		Unseen Client #4		Average
	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM
FedAvg [ 31 ]	35.57	88.68	34.00	85.26	32.52	76.71	36.86	91.44	32.89	81.35
FedProx [ 23 ]	36.69	91.16	34.30	86.39	32.58	72.13	39.13	94.87	34.74	85.52
FedNova [ 37 ]	35.45	88.83	34.23	86.29	32.39	69.20	35.40	89.99	34.38	83.58
MOON [ 22 ]	35.21	78.19	35.30	79.81	34.55	83.07	35.56	78.94	35.16	80.00
FedDG [ 28 ]	34.86	79.54	34.79	80.91	33.15	80.31	34.49	77.42	34.32	79.55

Table 3: Quantitative evaluation on unseen clients: PSNR and SSIM

Figure 3: Qualitative comparison of methods across different clients using the Transformer-based reconstruction network. Rows 1-5 show results for Clients #2, #3, #5, #6, and #7, respectively.

Component					Average
Pers. Domain	Proj. Cons.	Proto.	Anat.	Uncert.	PSNR	SSIM
		Aware	Aware	Weight	(dB)	(%)
Image	✗	✗	✗	✗	40.72	97.92
Image	✗	✓	✗	✗	40.98	97.98
Image	✗	✗	✓	✗	41.07	98.01
Image	✗	✓	✓	✗	41.12	98.03

Table 4: Ablation study on different components in ProFed.

Configuration	PSNR (dB)	SSIM (%)
Image loss only	41.16	98.05
+ Forward projection	41.37	98.09
+ Backward projection	41.29	98.07
+ Cycle consistency	41.33	98.08
+ All three ( $\lambda_{f}=1,\lambda_{b}=1,\lambda_{c}=1$ )	41.62	98.12

Table 5: Ablation study on individual projection losses

	Perf.	Arch. & Scale			Computation
Method	PSNR	Params	Params	Scale	Train	Infer	Extract	Comm.
		(K=8)	(K=100)
SCAN	43.65	1.5M	18.8M	$O(K)$	Base	Backbone	Per-ep.	Full
ProFed	44.83	3.0M	3.0M	$O(1)$	+Proj	Same	Offline	-14%
					(+0.46 dB)

Table 6: Computational Comparison

LDCT Denoising

Low-dose computed tomography (LDCT) denoising reduces radiation-induced image noise while preserving diagnostically important anatomy.

See all 20 papers

PatchDenoiser: Parameter-efficient multi-scale patch learning and fusion denoiser for Low-dose CT imaging
Feb 25, 2026Jitindra Fartiyal, Pedro Freire, Sergei K. Turitsyn +1LDCT Denoising Image Denoising
Aston Institute of Photonic Technologies, Aston University, Aston Triangle, Birmingham, B4 7ET, West Midlands, United Kingdom
FoundDiff: Foundational Diffusion Model for Generalizable Low-Dose CT Denoising
Aug 24, 2025Zhihao Chen, Qi Gao, Zilong Li +4Diffusion Models for Image Restoration Medical Imaging Foundation Models
Institute of Science and Technology for Brain-inspired Intelligence and MOE Frontiers Center for Brain Science, Fudan University, Shanghai 200433, China · Shanghai Key Lab of Intelligent Information Processing, School of Computer Science, Fudan University, Shanghai 200433, China · School of Cyber Science and Engineering, Sichuan University, Chengdu, Sichuan 610065, China +1
FedProIn: Mitigating Client Drift for Learnable Prototypes in Federated Medical Imaging
Jul 5, 2026Harsh Kumar, Tarun Kumar Garg, Vaanathi SundaresanFederated Learning Aggregation Non-IID Federated Learning
Department of Computational and Data Sciences, Indian Institute of Science Bengaluru, Karnataka 560012, India · Department of IISc Mathematics Initiative, Indian Institute of Science Bengaluru, Karnataka 560012, India

Feb 25, 2026J/K move · Enter open · S save

PatchDenoiser: Parameter-efficient multi-scale patch learning and fusion denoiser for Low-dose CT imaging

Jitindra Fartiyal, Pedro Freire, Sergei K. Turitsyn, Sergei G. Solovski

Aston Institute of Photonic Technologies, Aston University, Aston Triangle, Birmingham, B4 7ET, West Midlands, United Kingdom

LDCT Denoising Image Denoising Image Restoration

Open paper AskPDF arXiv

Low-dose CT images are essential for reducing radiation exposure in cancer screening, pediatric imaging, and longitudinal monitoring protocols, but their quality is often degraded by noise from low-dose acquisition, patient motion, or scanner limitations, affecting both clinical interpretation and downstream analysis. Traditional filtering approaches often over-smooth and lose fine anatomical details, while deep learning methods, including CNNs, GANs, and transformers, may struggle to preserve such details or require large, computationally expensive models, limiting clinical practicality. We propose PatchDenoiser, a lightweight, energy-efficient multi-scale patch-based denoising framework. It decomposes denoising into local texture extraction and global context aggregation, fused via a spatially aware patch fusion strategy. This design enables effective noise suppression while preserving fine structural and anatomical details. PatchDenoiser is ultra-lightweight, with far fewer parameters and lower computational complexity than CNN, GAN, and transformer based denoisers. On the 2016 Mayo Low-Dose CT dataset, PatchDenoiser consistently outperforms state-of-the-art CNN- and GAN-based methods in PSNR and SSIM. It is robust to variations in slice thickness, reconstruction kernels, and HU windows, generalizes across scanners without fine-tuning, and reduces parameters by ~9x and energy consumption per inference by ~27x compared with conventional CNN denoisers. PatchDenoiser thus provides a practical, scalable, and computationally efficient solution for medical image denoising, balancing performance, robustness, and clinical deployability.

Abstract

LDCT Denoising

Explore similar work

PatchDenoiser: Parameter-efficient multi-scale patch learning and fusion denoiser for Low-dose CT imaging

FoundDiff: Foundational Diffusion Model for Generalizable Low-Dose CT Denoising

FedProIn: Mitigating Client Drift for Learnable Prototypes in Federated Medical Imaging