[Paper Review] DenoMamba: A fused state-space model for low-dose CT denoising

Şaban Öztürk, Özge Duran+1 more|arXiv (Cornell University)|Sep 19, 2024

Medical Imaging Techniques and Applications7 citations

TL;DR

DenoMamba introduces a novel state-space modeling approach (FuseSSM blocks) to fuse spatial and channel context for LDCT denoising, achieving superior image quality over CNN/transformer baselines at similar complexity.

ABSTRACT

Low-dose computed tomography (LDCT) lower potential risks linked to radiation exposure while relying on advanced denoising algorithms to maintain diagnostic quality in reconstructed images. The reigning paradigm in LDCT denoising is based on neural network models that learn data-driven image priors to separate noise evoked by dose reduction from underlying tissue signals. Naturally, the fidelity of these priors depend on the model's ability to capture the broad range of contextual features evident in CT images. Earlier convolutional neural networks (CNN) are highly adept at efficiently capturing short-range spatial context, but their limited receptive fields reduce sensitivity to interactions over longer distances. Although transformers based on self-attention mechanisms have recently been posed to increase sensitivity to long-range context, they can suffer from suboptimal performance and efficiency due to elevated model complexity, particularly for high-resolution CT images. For high-quality restoration of LDCT images, here we introduce DenoMamba, a novel denoising method based on state-space modeling (SSM), that efficiently captures short- and long-range context in medical images. Following an hourglass architecture with encoder-decoder stages, DenoMamba employs a spatial SSM module to encode spatial context and a novel channel SSM module equipped with a secondary gated convolution network to encode latent features of channel context at each stage. Feature maps from the two modules are then consolidated with low-level input features via a convolution fusion module (CFM). Comprehensive experiments on LDCT datasets with 25\% and 10\% dose reduction demonstrate that DenoMamba outperforms state-of-the-art denoisers with average improvements of 1.4dB PSNR, 1.1% SSIM, and 1.6% RMSE in recovered image quality.

Motivation & Objective

• Motivate LDCT denoising to balance radiation dose reduction with diagnostic image quality.
• Propose a new SSM-based denoising architecture (DenoMamba) that captures short- and long-range context.
• Show that fused spatial-channel SSM blocks outperform CNNs and transformers with comparable complexity.
• Demonstrate robust denoising across 25% and 10% dose reductions on LDCT datasets.
• Provide an open-source implementation to foster reproducibility.

Proposed method

• Introduce FuseSSM blocks that combine a spatial SSM module and a channel SSM module.
• Use a secondary gated convolution network within the channel SSM to extract latent channel features.
• Fuse contextual representations with original inputs via a convolutional fusion module (CFM).
• Adopt an hourglass encoder-decoder architecture with long-range residual connections between encoder and decoder stages.
• Train with mean-squared error loss on LDCT-to-NDCT image pairs and evaluate with PSNR, SSIM, and RMSE.

Experimental results

Research questions

• RQ1 Can a fused spatial-channel state-space model improve LDCT denoising compared to CNN/transformer baselines at low-dose levels?
• RQ2 Does the proposed DenoMamba architecture maintain high spatial precision while capturing long-range context efficiently?
• RQ3 What is the impact of removing components (cha SSM, spa SSM, CFM, GCN, identity path) on denoising performance?
• RQ4 How does DenoMamba perform across 25% and 10% dose reductions and multiple reconstruction kernels?
• RQ5 Is the method code available for reproducibility?

Key findings

• DenoMamba achieves higher PSNR and SSIM and lower RMSE than competing methods on 25% dose LDCT datasets.
• Ablation studies show the importance of both spatial SSM and channel SSM modules and the GCN-based channel pathway for best performance.
• The proposed method demonstrates superior image quality across 1-mm and 3-mm slice thicknesses and various kernels.
• The model operates with comparable parameter counts to conventional CNNs while capturing long-range context with linear-like complexity.
• Code for DenoMamba is publicly available.