Generative 3D breast shape modeling and deep learning for x-ray scatter correction in dedicated breast CT: Towards 4D dynamic contrast-enhanced breast CT imaging

Background: Scatter correction is crucial to enhance image quality in x-ray imaging. Deep learning (DL) approaches have been proposed as alternatives to Monte Carlo (MC) simulations, offering comparable accuracy with faster processing times. Further research on DL approaches is required to include a wider range of breast shapes, and adapt scatter correction to the novel 4D dynamic contrast-enhanced dedicated breast CT (4D DCE-bCT) modality. Purpose: To develop a 3D generative breast shape model, providing a comprehensive representation of the female population, and enhance DL-based scatter correction for use in 4D DCE-bCT. Methods: 118 patient scans acquired with the bCT system at Radboudumc were segmented into the major breast tissue types (skin, adipose, fibroglandular tissue). The resulting patient-based breast phantoms were divided into training and validation sets, and augmented by random translation in the field of view. To accurately represent the variety of shapes observed in bCT images, a generative breast shape model was developed using a dataset of 491 patient scans from multiple clinical sites. The scans were represented by point clouds on a regular grid and then used for Principal Component Analysis (PCA). The model’s accuracy in capturing shape variance with linearly-independent vectors was assessed by computing the mean absolute error (MAE). Correlations between PCA parameters and breast characteristics (volume, chest-to-nipple distance, and radius at chest wall) were analyzed. The generated breast models were parameterized by selecting coefficients at the extremes of the PCA parameter space identified as the most influential on the overall shape. These models were converted into 3D representations by mesh reconstruction and voxelization and mapped with previously-segmented fibroglandular tissue through image registration. A subset of 104 newly-generated phantoms was added to the training and validation sets. MC simulations were performed to generate primary and scatter bCT images from each phantom. A previously-developed DL model tailored to 4D DCE-bCT was used for scatter estimation. The DL model was trained single projection-wise, providing thickness maps and breast locations as additional inputs, and tested on 10 phantoms generated from scans acquired with a different bCT system. The DL model’s performance was evaluated using the mean relative difference (MRD) and MAE. To evaluate the accuracy in the resulting reconstructed images, a full bCT acquisition was mimicked using MC simulations. MAE and structural similarity (SSIM) were computed between 100 slices of the DL-corrected and MC primary reconstructions. Results: The generative breast shape model achieved a median MAE of 0.6 mm compared to the scanned shapes. The PCA-generated breast phantoms effectively expanded the range of breast diameters included in the dataset. The DL model demonstrated mean MRD and MAE values of 0.27% (−0.88%, 1.60%) and 5.54% (4.63%, 6.11%) for the validation set, and −0.16% (−0.61%, 0.73%) and 5.42% (5.12%, 5.71%) for the external test set. MC-simulated reconstruction slices had an SSIM of 0.99995 and a MAE of 0.64% (0.45%, 0.80%). Conclusions: The DL model achieved satisfactory scatter estimation performance across a more comprehensive representation of the female population. The efficient processing time makes it suitable for routine clinical practice, and will enhance the accuracy of 4D DCE-bCT, allowing for accurate quantification of contrast-enhancement in tissues and lesions.