Window-Level Is a Strong Denoising Surrogate

CT image quality is heavily reliant on radiation dose, which causes a trade-off between dose and that affects the subsequent image-based diagnostic performance. However, high can be harmful to both patients operators. Several (deep learning-based) approaches have been attempted denoise low images. those require access large training sets, specifically full images for reference, often difficult obtain. Self-supervised learning an emerging alternative lowering reference data requirement facilitating unsupervised learning. Currently available self-supervised denoising works are either dependent foreign domains or pretexts not very task-relevant. To tackle aforementioned challenges, we propose novel approach, namely Self-Supervised Window-Leveling Image DeNoising (SSWL-IDN), leveraging innovative, task-relevant, simple, yet effective surrogate—prediction of window-leveled equivalent. SSWL-IDN leverages residual hybrid loss combining perceptual MSE, all incorporated in VAE framework. Our extensive (in- cross-domain) experimentation demonstrates effectiveness aggressive (abdomen chest) acquired at 5% level only (Code https://github.com/ayaanzhaque/SSWL-IDN).