No-Reference Denoising Of Low-Dose Ct Projections

Elvira Zainulina, Alexey Chernyavskiy, Dmitry V. Dylov

TL;DR

This paper introduces Noise2NoiseTD, a novel self-supervised deep learning method for denoising low-dose CT (LDCT) projections. It uniquely leverages the temporal redundancy between adjacent CT projections, eliminating the need for paired low- and high-dose data, and achieves performance comparable to supervised methods while outperforming existing self-supervised techniques.

What Problem Does This Paper Address?

Medical imaging, particularly computed tomography (CT), faces a critical challenge: reducing patient exposure to X-ray radiation. Low-dose CT (LDCT) is a promising solution, but it inherently introduces significant noise into the acquired projections. This noise degrades image quality, potentially impacting diagnostic accuracy and leading to artifacts in reconstructed CT slices. While supervised deep learning methods can effectively denoise these images, they typically require pairs of low- and high-dose scans of the exact same scene for training. Acquiring such paired data is impractical in clinical settings, as it doubles scan time and radiation dose. Existing self-supervised methods, while avoiding the need for paired data, often struggle with over-smoothing or introduce computational complexity. This paper aims to bridge this gap by developing a self-supervised denoising method that works with only noisy LDCT projections and can effectively utilize information from sequences of these projections.

Methods

The core innovation of this work is the Noise2NoiseTD approach, a self-supervised method that exploits the temporal correlation within a sequence of CT projections. The fundamental assumption is that while noise varies independently for each projection, the underlying anatomical structures remain consistent across a short sequence of projections taken from slightly different angles.

The method employs a U-Net-like architecture but with modifications: it avoids pooling layers to preserve spatial resolution and incorporates bidirectional Convolutional LSTM (Bi-ConvLSTM) units. These Bi-ConvLSTM units are crucial for capturing information from adjacent frames in both forward and backward temporal directions, allowing the network to learn how structures change subtly across projections. A CNN backbone extracts features from individual frames, which are then processed by the Bi-ConvLSTM layers. The outputs from these layers are aggregated, with the network trained to predict the denoised middle projection of the sequence.

Crucially, Noise2NoiseTD is self-supervised, meaning it doesn't require ground truth clean images. Instead, it uses a no-reference loss function derived from a probabilistic model of noise in CT projections. This loss function is designed to account for a mixed Poisson-Gaussian noise distribution, which the authors found to be more accurate for LDCT than purely Gaussian or Poisson models. The network learns to estimate the parameters of this noise distribution alongside the denoising task.

Datasets

The experiments were conducted using the LDCT dataset published by the Mayo Clinic. This dataset provides both full-dose and simulated low-dose (25% of routine dose) CT projections. The authors trained two independent models: one for abdominal projections and another for head projections, as the noise characteristics and dose levels differ between these modalities. For training and validation, they used randomly selected projections from seven patients (five for head, two for abdomen). Test sets were also curated from these patients, with additional testing performed on data from two other patients to assess generalizability.

Key Insights and Ideas

  • Temporal Redundancy is Key: The most significant insight is that the subtle changes in CT projections due to rotation are sufficient to distinguish anatomical structures from noise, especially when considering a sequence of projections. This temporal information is largely untapped by previous self-supervised methods that treat each projection independently.
  • Bi-ConvLSTM for Temporal Context: The use of Bi-ConvLSTM is a clever way to explicitly model and leverage the temporal dependencies between projections, providing a richer context for denoising than simple frame-wise processing.
  • Probabilistic Noise Modeling for Self-Supervision: Developing a no-reference loss function that accurately models the mixed Poisson-Gaussian noise distribution is critical for effective self-supervised learning in this domain. This allows the network to learn without explicit ground truth.
  • Performance Without Ground Truth: The paper demonstrates that self-supervised methods can achieve performance remarkably close to supervised methods, which is a major step towards practical clinical deployment where ground truth is scarce.

Key Results

  • Competitive Performance: Noise2NoiseTD achieved denoising accuracy comparable to supervised methods, as measured by SSIM (Structural Similarity Index Measure) and PSNR (Peak Signal-to-Noise Ratio).
  • Superior to Existing Self-Supervised Methods: The proposed method outperformed modern self-supervised denoising techniques like Noise2Void-4R, particularly in preserving edges and fine details in reconstructed CT scans.
  • Visual Quality: Qualitative results showed that Noise2NoiseTD produced less over-smoothed images compared to Noise2Void-4R, even when the latter used multiple adjacent projections. The denoised images were visually comparable to those obtained by supervised methods.
  • Generalizability: The model showed good generalizability when tested on data from different patients than those used for training.

Limitations and Future Work

  • Simulated Data: The experiments were conducted on simulated low-dose data. The authors acknowledge the need to validate the method on real-world LDCT data.
  • Noise Parameter Estimation: While the method estimates noise parameters, further refinement of the learning process to improve noise approximation and parameter estimation is planned.
  • Radiologist Validation: The authors suggest that image quality metrics may not perfectly align with radiologists' opinions. Future work should involve qualitative comparisons with domain experts.
  • Broader Applicability: The method's core principles are not specific to CT and could be extended to denoise sequences of other medical imaging modalities, such as dynamic PET, MR, spectral CT, and ultrasonograms.

Summary updated

Abstract

Low-dose computed tomography (LDCT) became a clear trend in radiology with an aspiration to refrain from delivering excessive X-ray radiation to the patients. The reduction of the radiation dose decreases the risks to the patients but raises the noise level, affecting the quality of the images and their ultimate diagnostic value. One mitigation option is to consider pairs of low-dose and high-dose CT projections to train a denoising model using deep learning algorithms; however, such pairs are rarely available in practice. In this paper, we present a new self-supervised method for CT denoising. Unlike existing self-supervised approaches, the proposed method requires only noisy CT projections and exploits the connections between adjacent images. The experiments carried out on an LDCT dataset demonstrate that our method is almost as accurate as the supervised approach, while also outperforming several modern self-supervised denoising methods.
BibTeX
@inproceedings{zainulina2021no,
  author = {Elvira Zainulina and Alexey Chernyavskiy and Dmitry V. Dylov},
  title = {{No-Reference Denoising Of Low-Dose Ct Projections}},
  booktitle = {ISBI},
  year = {2021},
  url = {https://openreview.net/forum?id=ohAlOoNv7W},
}
Figure