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

TL;DR

This paper tackles privacy-preserving LDCT denoising under severe data heterogeneity by introducing SCAN-PhysFed, a dual-level physics-informed federated learning framework. It couples anatomy-level prompts derived from an LLM-generated radiology report with scanning-level prompts from protocol data via two hypernetworks, enabling patient- and protocol-specific feature modulation. A novel Protocol Vector-Quantization Strategy (PVQS) ensures robust performance on unseen protocols, while an orthogonality loss promotes discriminative protocol codes. Empirical results on public LDCT datasets demonstrate superior denoising performance and stability across diverse protocols and backbones, highlighting the approach's potential for practical, privacy-preserving CT reconstruction in heterogeneous clinical settings.

Abstract

Reducing radiation doses benefits patients, however, the resultant low-dose computed tomography (LDCT) images often suffer from clinically unacceptable noise and artifacts. While deep learning (DL) shows promise in LDCT reconstruction, it requires large-scale data collection from multiple clients, raising privacy concerns. Federated learning (FL) has been introduced to address these privacy concerns; however, current methods are typically tailored to specific scanning protocols, which limits their generalizability and makes them less effective for unseen protocols. To address these issues, we propose SCAN-PhysFed, a novel SCanning- and ANatomy-level personalized Physics-Driven Federated learning paradigm for LDCT reconstruction. Since the noise distribution in LDCT data is closely tied to scanning protocols and anatomical structures being scanned, we design a dual-level physics-informed way to address these challenges. Specifically, we incorporate physical and anatomical prompts into our physics-informed hypernetworks to capture scanning- and anatomy-specific information, enabling dual-level physics-driven personalization of imaging features. These prompts are derived from the scanning protocol and the radiology report generated by a medical large language model (MLLM), respectively. Subsequently, client-specific decoders project these dual-level personalized imaging features back into the image domain. Besides, to tackle the challenge of unseen data, we introduce a novel protocol vector-quantization strategy (PVQS), which ensures consistent performance across new clients by quantifying the unseen scanning code as one of the codes in the scanning codebook. Extensive experimental results demonstrate the superior performance of SCAN-PhysFed on public datasets.

Figures (8)

• Figure 1: Illustration of different radiation reduction strategies. "Sparse-View" and "Low-Dose" represent the scanning with sparse views and low incident photons, respectively.
• Figure 2: The overall learning paradigm of our proposed SCAN-PhysFed.
• Figure 3: Simulated examples under different protocols. (a)-(h) show examples from training clients with known protocols, while (i)-(l) present examples from unseen protocols.
• Figure 4: 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. The display window for the first row is [-1024, 200] HU, while for the other rows, it is [-160, 240] HU.
• Figure 5: Boxplots of the average results across all clients. (a) and (b) represent PSNR and SSIM, respectively.