[Paper Review] Projection Guided Personalized Federated Learning for Low Dose CT Denoising

Anas Zafar, Muhammad Waqas+3 more|arXiv (Cornell University)|Mar 12, 2026

Medical Imaging Techniques and Applications0 citations

TL;DR

ProFed performs projection-space dual-level personalization for federated LDCT denoising, using sinogram measurements for physics-based supervision and uncertainty-weighted aggregation to better separate scanner noise from patient anatomy, outperforming 11 baselines.

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.

Motivation & Objective

• Motivate separation of protocol-dependent noise from patient anatomy in LDCT reconstruction.
• Develop a federated framework that personalizes in projection space to handle heterogeneity across scanners.
• Introduce dual adaptation (anatomy-aware and protocol-aware networks) with projection-domain consistency losses.
• Incorporate uncertainty-guided client weighting to enhance robust aggregation across institutions.

Proposed method

• Operate in projection space using differentiable Radon transform and backprojection to supervise reconstruction with CT measurements.
• Dual adaptation: anatomy-aware network guided by image and text-derived anatomy features, and protocol-aware (scanner) adaptation via LoRA-based modulation.
• Multi-constraint projection losses: forward projection, backward projection, and cycle consistency losses with sinograms.
• Uncertainty-guided aggregation: Monte Carlo dropout to estimate client confidence and weight contributions during aggregation.
• Loss composition combines image-domain reconstruction, heteroscedastic noise modeling, and projection-domain consistency.
• Training uses FedAvg-like federated rounds with an O(1) parameter scaling and offline anatomy feature extraction.

Experimental results

Research questions

• RQ1 Can projection-domain (sinogram-space) supervision better separate scanner noise from anatomy compared to image-space methods in LDCT federated learning?
• RQ2 Does dual adaptation (anatomy-aware and protocol-aware) with projection consistency improve reconstruction quality across heterogeneous protocols?
• RQ3 Is uncertainty-guided aggregation effective for robust federated learning across multiple institutions with varying protocols?
• RQ4 What is the impact of projection-based losses (forward, backward, cycle) on LDCT denoising performance?
• RQ5 How well does ProFed generalize to unseen patients and unseen protocols?

Key findings

• ProFed achieves 42.56 dB PSNR with CNN backbones on Mayo Clinic 2016, surpassing 11 federated baselines.
• ProFed achieves 44.83 dB PSNR with Transformer backbones, outperforming baselines by up to +1.42 dB over SCAN-PhysFed.
• On unseen clients, ProFed attains 39.17 dB PSNR, showing robust generalization to new protocols and patients.
• Uncertainty-guided aggregation improves effective federated learning, with projection-domain supervision providing strong gains over image-space methods.
• Ablation shows full projection-guided dual adaptation yields the best performance, with about +0.9 to +1.4 dB gains from adding projection losses and dual adaptation components.
• In terms of scalability, ProFed maintains O(1) parameter growth with federation size and reduces communication overhead by 14% via offline anatomy feature extraction.