A deep dictionary network-based foundation model for ultra-low-dose CT denoising

Baoshun Shi; Shuangyi Yang; Ke Jiang; Bin Zhu; Zhanli Hu; Huazhu Fu

eess.IVSep 11, 2026

A deep dictionary network-based foundation model for ultra-low-dose CT denoising

Authors: Baoshun Shi, Shuangyi Yang, Ke Jiang, Bin Zhu, Zhanli Hu, Huazhu Fu

Organizations: School of Information Science and Engineering, Yanshan University, Qinhuangdao 066004, Hebei, China · Hebei Key Laboratory of Information Transmission and Signal Processing, Yanshan University, Qinhuangdao 066004, Hebei, China · Department of Orthopedics, Beijing Friendship Hospital, Capital Medical University, Beijing, China · Lauterbur Research Center for Biomedical Imaging, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, Shenzhen 518055, China ·

Abstract

Ultra-low-dose computed tomography (ULDCT) reduces radiation exposure but suffers from severe noise that degrades diagnostic image quality. Existing deep learning-based denoising methods are typically trained in an organ-specific fashion, resulting in limited generalization across heterogeneous multi?organ imaging scenarios. Foundation models present a promising all-in-one paradigm for unified multi-organ denoising. However, their architectures suffer from poor interpretability and rely on heuristic training strategies. To address these limitations, we propose an architecture?interpretable foundation model based on the deep dictionary network (DDN) for unified multi-organ ULDCT denoising. Inspired by multilayer sparse representation theory, DDN cascades convolutional sparse coding layers with iterative soft-thresholding, providing inherent architectural interpretability. Furthermore, a dynamic dictionary module and a threshold generation module are embedded within each layer to enhance representation ability. We conduct DDN pre-training on more than one million multi-organ normal-dose CT images by recovering clean images from Gaussian-noised inputs. Sparse regularization is additionally imposed on latent feature representations, guiding the network to learn compact and noise-robust priors. The complete architecture is jointly fine-tuned on multi-organ ULDCT datasets, enabling a single unified model to perform denoising across diverse anatomical regions. Extensive experiments validate that our proposed method achieves state-of-the?art performance and consistently surpasses competing ULDCT methods across all mul

Figures & tables

Fig. 1: Overview of the proposed DDN-based pre-training framework and few-shot denoising performance. (a) DDN is pretrained on more than one million multi-organ CT images from ten datasets and then combined with a decoder for multi-organ ULDCT denoising, cross-organ generalization, and few-shot learning. (b) Average PSNR and SSIM comparison of different methods on test datasets under the few-shot setting, where only 20 training images are used for each organ. DDN achieves the highest PSNR and SSIM among the compared methods, indicating better denoising performance under limited training data.

Fig. 2: Architecture of the proposed DDN. DDN cascades convolutional sparse coding layers with iterative updates, where each layer incorporates input-dependent DDM ( $\bm{D}_{j}$ and $\tilde{\bm{D}}_{j}$ ) and adaptive soft-thresholding. The TGM integrates local, large, non-local, and global features to produce a spatially adaptive threshold map shared across the iterative steps within each layer.

Method	Head	Colon	Hepatic Vessel	Lung	Spleen	Average
LDCT Models
FBP	28.02/0.8838/0.0398	31.30/0.8836/0.0272	30.98/0.8775/0.0284	31.30/0.8717/0.0273	30.76/0.8677/0.0290	30.47/0.8769/0.0303
REDCNN [ 3 ]	41.31/0.9722/0.0086	43.43/0.9705/0.0068	40.24/0.9521/0.0098	40.78/0.9476/0.0091	40.86/0.9555/0.0091	41.32/0.9596/0.0087
CNN10 [ 4 ]	41.00/0.9751/0.0090	42.63/0.9751/0.0074	39.69/0.9519/0.0104	40.16/0.9491/0.0098	39.84/0.9549/0.0102	40.67/0.9612/0.0094
EDCNN [ 5 ]	39.05/0.9613/0.0112	41.51/0.9599/0.0084	38.51/0.9310/0.0119	39.23/0.9372/0.0109	38.66/0.9293/0.0117	39.39/0.9437/0.0108
FoundDiff [ 21 ]	42.48/0.9837/0.0076	43.42/0.9811/0.0068	42.04/0.9717/0.0080	41.05/0.9596/0.0089	41.31/0.9702/0.0086	42.06/0.9733/0.0080

TABLE I: Quantitative evaluations for different ULDCT denoising methods on the test datasets. We report the average [PSNR (dB)

\uparrow

/SSIM

\uparrow

/RMSE

\downarrow

] values for each organ. The best results are highlighted in red and the second-best results are blue .

Fig. 3: Visual comparison of different ULDCT denoising methods on a representative lung case from the internal datasets. Images are displayed using an HU window of [50, 500]. Green boxes indicate enlarged regions for detailed comparison of noise suppression and anatomical structure preservation.

Fig. 4: Visual comparison of different ULDCT denoising methods on a representative kidney case from the external datasets. Images are displayed using an HU window of [40, 350]. Green boxes indicate enlarged regions for detailed comparison of cross-organ denoising performance.

Fig. 5: Visual comparison of different ULDCT denoising methods on real pork with bone data. Green boxes indicate enlarged regions for detailed comparison of noise suppression and structural preservation under practical acquisition conditions.

LDCT Models
Method	Kidney	Pancreas	Average
FBP	30.47/0.8964	31.11/0.8776	30.79/0.8870
REDCNN	41.14/0.9733	42.48/0.9209	41.81/0.9471
CNN10	41.01/0.9673	42.21/0.9671	41.61/0.9672
EDCNN	40.01/0.9576	41.64/0.9613	40.83/0.9595
FoundDiff	40.93/0.9584	41.93/0.8653	41.43/0.9119

TABLE II: Results on external datasets for kidney and pancreas. We report the average [PSNR (dB)

\uparrow

/SSIM

\uparrow

] values for each organ. The best and second-best results are highlighted in red and blue , respectively.

Fig. 6: Effect of decoder design and foundation model initialization on ULDCT denoising. The boxplots show the PSNR and SSIM distributions over the five organ test sets. ”2D+Conv41” denotes Conv41 with two additional dictionary layers, while ”w/o FM” and ”with FM” denote training without and with the pretrained DDN foundation model, respectively.

Fig. 7: Visual comparison of different ULDCT denoising methods on a representative colon case under the few-shot setting. The enlarged regions are used for detailed comparison of local anatomical structures.

Fig. 8: Effect of decoder design and foundation model initialization under the few-shot setting. PSNR and SSIM are reported for three decoder configurations with and without foundation model initialization, while the red annotations indicate the corresponding performance gains introduced by FM.

Method	PSNR (dB)	SSIM	RMSE
DDN w/ 1 dictionary layer	42.49	0.9752	0.0076
DDN w/ 2 dictionary layers	42.75	0.9762	0.0074
w/o local in TGM	42.83	0.9768	0.0073
w/o large in TGM	42.80	0.9762	0.0073
w/o non-local in TGM	41.27	0.9694	0.0088
w/o global in TGM	42.75	0.9762	0.0074

TABLE III: Ablation studies on key components of the proposed method. The reported metrics are averaged over the test datasets. The best results are shown in bold.

LDCT Denoising

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

See all 20 papers

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
Unsupervised Denoising of Real Clinical Low Dose Liver CT with Perceptual Attention Networks
May 1, 2026Zhilin Guan, Wei ZhangComputed Tomography LDCT Denoising
Department of Computing Harbin Institute of Technology Harbin, China
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

Aug 24, 2025J/K move · Enter open · S save

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

Zhihao Chen, Qi Gao, Zilong Li, Junping Zhang+3

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

Diffusion Models for Image Restoration Medical Imaging Foundation Models Computed Tomography LDCT Denoising

Open paper AskPDF arXiv

Low-dose computed tomography (CT) denoising is crucial for reduced radiation exposure while ensuring diagnostically acceptable image quality. Despite significant advancements driven by deep learning (DL) in recent years, existing DL-based methods, typically trained on a specific dose level and anatomical region, struggle to handle diverse noise characteristics and anatomical heterogeneity during varied scanning conditions, limiting their generalizability and robustness in clinical scenarios. In this paper, we propose FoundDiff, a foundational diffusion model for unified and generalizable LDCT denoising across various dose levels and anatomical regions. FoundDiff employs a two-stage strategy: (i) dose-anatomy perception and (ii) adaptive denoising. First, we develop a dose- and anatomy-aware contrastive language-image pre-training model (DA-CLIP) to achieve robust dose and anatomy perception by leveraging specialized contrastive learning strategies to learn continuous representations that quantify ordinal dose variations and identify salient anatomical regions. Second, we design a dose- and anatomy-aware diffusion model (DA-Diff) to perform adaptive and generalizable denoising by synergistically integrating the learned dose and anatomy embeddings from DA-CLIP into diffusion process via a novel dose and anatomy conditional block (DACB) based on Mamba. Extensive experiments on a large simulated multi-dose CT dataset spanning three anatomical regions, together with cross-dataset evaluations on Mayo-2016, CQ500, and piglet datasets, demonstrate superior denoising performance and strong generalization to unseen dose levels and anatomical regions. The codes and models are available at https: //github.com/hao1635/FoundDiff.

Abstract

LDCT Denoising

Explore similar work

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

Unsupervised Denoising of Real Clinical Low Dose Liver CT with Perceptual Attention Networks

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