Introduction
Dynamic Computed Tomography Perfusion (CTP) is a cornerstone in the diagnosis of acute ischemic strokes, providing critical maps of cerebral blood flow by tracking a contrast agent through the brain. However, the repeated scanning required for these temporal maps exposes patients to significant radiation. While reducing the radiation dose is desirable, low-dose CT imaging introduces substantial "salt-and-pepper" noise and artifacts. These distortions can degrade the quality of hemodynamic maps, potentially leading to misdiagnosis. Traditional denoising methods often blur structural details, and standard deep learning approaches typically require high-dose "ground truth" images, which are ethically and logistically difficult to obtain in clinical practice.
Methods
This research introduces a self-supervised deep learning framework that eliminates the need for clean reference scans by adapting the Noise2Noise principle for temporal medical data. The core strategy leverages the high correlation between adjacent video frames in a CTP sequence.
Self-Supervised Temporal Mapping
Instead of training a network to map a noisy image to a clean one, the model is trained to map a specific frame $x(t)$ to an estimation derived from its temporal neighbors, $x(t-1)$ and $x(t+1)$. Because the noise in these frames is independent while the underlying anatomy is consistent, the network learns to extract the common signal.
Technical Refinements
To handle the rapid intensity changes during contrast agent transit, the authors implemented several key features:
- Linear Correction: A correction factor ($\kappa$) is applied to neighbors to better estimate the current frame's intensity, preventing bias during peak contrast flow.
- Bias Compensation Loss: A specific loss term ensures that the low-frequency structural components of the denoised image remain consistent with the original noisy input, preventing over-smoothing.
- Efficient Backbone: The system utilizes a 2D UNet architecture, ensuring rapid inference times suitable for emergency clinical environments.
Datasets and Evaluation
The model was validated using two distinct datasets:
- Simulation Data: Open-source phantoms were used to create ground-truth benchmarks for infarct cores and penumbra areas, allowing for precise quantitative error measurement.
- Real-World Data: The model was tested on clinical scans from 20 patients in the 2018 ISLES challenge, evaluating its performance against actual clinical noise and artifacts.
Results
In simulation trials, the proposed self-supervised method significantly outperformed conventional filters such as Gaussian, TIPS, and TTV. Notably, it achieved performance levels nearly identical to supervised models that had access to clean ground-truth data.
On the real-world ISLES dataset, the self-supervised model achieved a Contrast-to-Noise Ratio (CNR) of 3.29, surpassing both the supervised model (2.57) and traditional filtering techniques. Furthermore, the model is computationally efficient, processing a standard CTP slice in approximately 1.5 seconds on a GPU, which is vital for time-sensitive stroke interventions.
Discussion and Conclusion
The "neighbor as label" strategy effectively bypasses the need for high-dose training data by utilizing the temporal redundancy inherent in dynamic imaging. By training directly on real clinical data, the model avoids the "domain shift" issues common in supervised models trained on synthetic data.
While the model assumes zero-mean noise and is sensitive to extreme patient motion, it represents a significant step toward safer, low-dose diagnostic imaging. Future work may involve extending this framework to other dynamic modalities, such as myocardial CT perfusion or dynamic PET scans, and incorporating recurrent neural networks to leverage longer temporal sequences.
Summary updated