Research map / Federated learning and privacy / Personalized and heterogeneous FL

Patient-Level Anatomy Meets Scanning-Level Physics: Personalized Federated Low-Dose CT Denoising Empowered by Large Language Model

CVPR 2025 · Federated learning and privacy › Personalized and heterogeneous FL · approach: Sparse and collaboration-graph personalization · 18 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SCAN-PhysFed, a personalized federated LDCT denoising framework that uses scanning-protocol and LLM-generated radiology-report prompts in hypernetworks plus a protocol vector-quantization strategy for unseen clients.

Problem

Federated low-dose CT denoising suffers from data heterogeneity across scanning protocols and anatomies, and existing FL methods generalize poorly to unseen scanning protocols.

Approach

Dual-level physics-informed personalization: a scanning-informed hypernetwork (from protocol parameters) and an anatomy-informed hypernetwork (from a medical LLM's radiology-report features) modulate shared imaging features, with client-specific decoders; unseen protocols are mapped to the nearest codebook entry (PVQS) and use that client's decoder.

Key results

• SCAN-PhysFed with RED-CNN achieves average PSNR 41.14 dB and SSIM 97.62% across 8 clients, versus 40.05 dB / 95.79% for the next best (FedFDD).
• With Uformer backbone, SCAN-PhysFed achieves average PSNR 43.65 dB and SSIM 97.03%, outperforming FedFDD (41.84 dB).
• On four unseen clients, SCAN-PhysFed reaches average PSNR 38.55 dB and SSIM 93.04%, above FedFDD (37.27 dB / 89.87%).

Measured effects reported in the paper

Surprising findings

• Personalized FL methods degrade more than generic FL methods on unseen clients, since local optimization underfits other distributions.

Limitations

• Evaluated only on simulated protocols from a single public dataset with 10 patients.
• Personalized decoders are client-specific, and the unseen-client baseline averages local models, which is a vanilla design.

Experimental setup

Datasets: 2016 NIH-AAPM-Mayo Clinic Low-Dose CT Grand Challenge

Benchmarks: 2016 NIH-AAPM-Mayo Clinic Low-Dose CT Grand Challenge

Metrics: PSNR SSIM

Baselines: FedAvg FedProx FedNova MOON FedDG FedKD FedPer FedBN FedMRI HyperFed FedFDD

Builds on: FedAvg HyperFed FedFDD RED-CNN miniGPT-Med

Closest papers by meaning

• Large-scale and Fine-grained Vision-language Pre-training for Enhanced CT Image Understanding (ICLR 2025 · similarity 0.85)
• FedVCK: Non-IID Robust and Communication-Efficient Federated Learning via Valuable Condensed Knowledge for Medical Image Analysis (AAAI 2025 · similarity 0.83)
• Federated Modality-Specific Encoders and Multimodal Anchors for Personalized Brain Tumor Segmentation (AAAI 2024 · similarity 0.83)
• Better Tokens for Better 3D: Advancing Vision-Language Modeling in 3D Medical Imaging (NeurIPS 2025 · similarity 0.83)
• Knowledge Transfer Scaling Laws for 3D Medical Imaging (NeurIPS 2026 · similarity 0.83)
• TumorChain: Interleaved Multimodal Chain-of-Thought Reasoning for Traceable Clinical Tumor Analysis (ICLR 2026 · similarity 0.82)