Self-Supervised Denoising of Low-Dose EPID Portal Images with Edge Preservation

Low-dose megavoltage (MV) portal imaging using Electronic Portal Imaging Devices (EPID) is essential for treatment verification and Quality Assurance (QA) but is limited by increased quantum noise and reduced edge blur. This blurring hinders accurate verification of the isocenter, field edges, and Multi-Leaf Collimator (MLC) boundaries. This study investigates whether self-supervised deep learning can effectively denoise low-dose EPID images while preserving edge sharpness and geometric fidelity. A self-supervised blind-spot denoising network was trained on unpaired EPID frames using 64–128-pixel patches, learning directly from noisy data without paired low/high-dose images. Model performance was assessed using Signal-to-Noise Ratio (SNR), edge contrast, and Modulation Transfer Function (MTF). Results demonstrate substantial noise reduction with preserved edge detail, outperforming classical filters and supervised methods. The proposed approach enhances the usability of low-dose portal imaging for routine QA and shows promise for broader clinical application following further validation.