Evaluation of deep-learning-based scatter correction in x-ray breast imaging: across image domains and downsampling ratios

Scatter radiation in X-ray breast imaging degrades contrast, hindering lesion detectability. As scatter signals are primarily low-frequency, downsampling is commonly applied in deep-learning-based correction methods. We investigated scatter estimation networks across image domains and downsampling ratios using VICTRE and MC-GPU simulations. Networks were trained in the inverted intensity and log-transformed domains with downsampling ratios of 2x2, 8x8, and 16x16. Performance was evaluated on scatter-corrected images restored to the original resolution. A U-Net model with residual connections was trained using mean absolute error (MAE). Mean squared error (MSE) and structural similarity index (SSIM) were calculated. Log-transformed training resulted in higher pixel accuracy (lower MSE), while the inverted intensity domain achieved higher SSIM, indicating better structural preservation. We suggest log-transformed domain training for tasks requiring pixel accuracy, whereas the inverted intensity domain is more suitable for preserving details, which is crucial for detecting anomalies in X-ray breast imaging.