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 can lead to contrast degradation, potentially hindering lesion detectability. To mitigate the issue, deep-learning-based methods for scatter correction have been widely studied. Since scatter signals are mainly low-frequency, downsampling techniques are commonly applied. In this study, we investigated the performance of scatter estimation networks under different image domains and downsampling ratios. Data was obtained using VICTRE and MC-GPU simulation. We trained networks on images in two domains: the inverted intensity domain and the log-transformed domain, with downsampling ratios of 2×2, 8×8, and 16×16. The performance metrics were evaluated on scatter-corrected images in the log-transformed domain, which were restored to the original resolution. A U-Net model with residual connections was trained for scatter estimation, and a mean absolute error (MAE) was used as the loss function. We calculated mean squared error (MSE) and structural similarity index (SSIM) for quantitative analysis. We found that training in the intensity domain with 2×2 downsampling was ineffective. Meanwhile, for other downsampling ratios, training in the log-transformed domain resulted in lower MSE, indicating higher pixel value accuracy. The network trained in the inverted intensity domain achieved higher SSIM values, which suggests better structural preservation. Qualitative comparisons indicated that small objects like microcalcifications were more clearly visualized in the intensity domain training. We suggest that log-transformed domain training is preferable when pixel accuracy is crucial, whereas the inverted intensity domain may be more suitable for preserving structural details, which is important for detecting anomalies in x-ray breast imaging.