.png)
April 22, 2025
Reducing image noise is crucial in low-dose computed tomography (LDCT) to improve diagnostic accuracy while minimizing radiation exposure for patients. Traditional noise reduction methods based on supervised learning require large amounts of training data with image pairs of noisy and noise-free acquisitions. However, obtaining such datasets is time-consuming and expensive. Self-supervised approaches offer a promising alternative, as they do not require such pairs. However, many of these methods require multiple noisy acquisitions of the same object and rely on complex neural networks whose functionality is often difficult to understand.
A new study presents an innovative approach to noise reduction in LDCT: Filter2Noise (F2N). This method utilizes self-supervised learning and requires only a single noisy acquisition for training. At the core of F2N is a so-called Attention-Guided Bilateral Filter. This filter is controlled by a small neural network that calculates the optimal filter parameters for each input. The parameters control the spatial and color smoothing of the image and can be visualized and adjusted after training. This allows for user-specific noise reduction in specific areas of the image that are relevant for diagnosis.
To enable training with only one image, F2N uses a novel downsampling-shuffle strategy in combination with a special loss function. This extends the concept of Noise2Noise to single images and takes into account the spatial correlation of the noise. Compared to existing self-supervised single-image methods, such as ZS-N2N, F2N achieves significantly better results in terms of Peak Signal-to-Noise Ratio (PSNR). Furthermore, F2N offers more transparency and user control and is very efficient due to the small number of parameters in the neural network.
The interpretability and the possibility of user-specific adaptation make F2N particularly attractive for medical applications. Doctors can adjust the filter parameters after training to adapt the noise reduction to specific diagnostic requirements and thus optimize the image quality in critical areas. Future research could focus on extending F2N to other imaging modalities and integrating it into clinical workflows. The development of robust and interpretable AI solutions like F2N contributes to improving diagnostic capabilities in medicine while reducing radiation exposure for patients.
