Simulation-driven training of vision transformers enables metal artifact reduction of highly truncated CBCT scans.

BACKGROUND: Due to the high attenuation of metals, severe artifacts occur in cone beam computed tomography (CBCT). The metal segmentation in CBCT projections usually serves as a prerequisite for metal artifact reduction (MAR) algorithms. PURPOSE: The occurrence of truncation caused by the limited detector size leads to the incomplete acquisition of metal masks from the threshold-based method in CBCT volume. Therefore, segmenting metal directly in CBCT projections is pursued in this work. METHODS: Since the generation of high quality clinical training data is a constant challenge, this study proposes to generate simulated digital radiographs (data I) based on real CT data combined with self-designed computer aided design (CAD) implants. In addition to the simulated projections generated from 3D volumes, 2D x-ray images combined with projections of implants serve as the complementary data set (data II) to improve the network performance. In this work, SwinConvUNet consisting of shift window (Swin) vision transformers (ViTs) with patch merging as encoder is proposed for metal segmentation. RESULTS: The model's performance is evaluated on accurately labeled test datasets obtained from cadaver scans as well as the unlabeled clinical projections. When trained on the data I only, the convolutional neural network (CNN) encoder-based networks UNet and TransUNet achieve only limited performance on the cadaver test data, with an average dice score of 0.821 and 0.850. After using both data II and data I during training, the average dice scores for the two models increase to 0.906 and 0.919, respectively. By replacing the CNN encoder with Swin transformer, the proposed SwinConvUNet reaches an average dice score of 0.933 for cadaver projections when only trained on the data I. Furthermore, SwinConvUNet has the largest average dice score of 0.953 for cadaver projections when trained on the combined data set. CONCLUSIONS: Our experiments quantitatively demonstrate the effectiveness of the combination of the projections simulated under two pathways for network training. Besides, the proposed SwinConvUNet trained on the simulated projections performs state-of-the-art, robust metal segmentation as demonstrated on experiments on cadaver and clinical data sets. With the accurate segmentations from the proposed model, MAR can be conducted even for highly truncated CBCT scans.

A deep learning approach to estimate x-ray scatter in digital breast tomosynthesis: From phantom models to clinical applications.

BACKGROUND: Digital breast tomosynthesis (DBT) has gained popularity as breast imaging modality due to its pseudo-3D reconstruction and improved accuracy compared to digital mammography. However, DBT faces challenges in image quality and quantitative accuracy due to scatter radiation. Recent advancements in deep learning (DL) have shown promise in using fast convolutional neural networks for scatter correction, achieving comparable results to Monte Carlo (MC) simulations. PURPOSE: To predict the scatter radiation signal in DBT projections within clinically-acceptable times and using only clinically-available data, such as compressed breast thickness and acquisition angle. METHODS: MC simulations to obtain scatter estimates were generated from two types of digital breast phantoms. One set consisted of 600 realistically-shaped homogeneous breast phantoms for initial DL training. The other set was composed of 80 anthropomorphic phantoms, containing realistic internal tissue texture, aimed at fine tuning the DL model for clinical applications. The MC simulations generated scatter and primary maps per projection angle for a wide-angle DBT system. Both datasets were used to train (using 7680 projections from homogeneous phantoms), validate (using 960 and 192 projections from the homogeneous and anthropomorphic phantoms, respectively), and test (using 960 and 48 projections from the homogeneous and anthropomorphic phantoms, respectively) the DL model. The DL output was compared to the corresponding MC ground truth using both quantitative and qualitative metrics, such as mean relative and mean absolute relative differences (MRD and MARD), and to previously-published scatter-to-primary (SPR) ratios for similar breast phantoms. The scatter corrected DBT reconstructions were evaluated by analyzing the obtained linear attenuation values and by visual assessment of corrected projections in a clinical dataset. The time required for training and prediction per projection, as well as the time it takes to produce scatter-corrected projection images, were also tracked. RESULTS: The quantitative comparison between DL scatter predictions and MC simulations showed a median MRD of 0.05% (interquartile range (IQR), -0.04% to 0.13%) and a median MARD of 1.32% (IQR, 0.98% to 1.85%) for homogeneous phantom projections and a median MRD of -0.21% (IQR, -0.35% to -0.07%) and a median MARD of 1.43% (IQR, 1.32% to 1.66%) for the anthropomorphic phantoms. The SPRs for different breast thicknesses and at different projection angles were within ± 15% of the previously-published ranges. The visual assessment showed good prediction capabilities of the DL model with a close match between MC and DL scatter estimates, as well as between DL-based scatter corrected and anti-scatter grid corrected cases. The scatter correction improved the accuracy of the reconstructed linear attenuation of adipose tissue, reducing the error from -16% and -11% to -2.3% and 4.4% for an anthropomorphic digital phantom and clinical case with similar breast thickness, respectively. The DL model training took 40 min and prediction of a single projection took less than 0.01 s. Generating scatter corrected images took 0.03 s per projection for clinical exams and 0.16 s for one entire projection set. CONCLUSIONS: This DL-based method for estimating the scatter signal in DBT projections is fast and accurate, paving the way for future quantitative applications.

Generative compressed breast shape model for digital mammography and digital breast tomosynthesis.

BACKGROUND: Modelling of the 3D breast shape under compression is of interest when optimizing image processing and reconstruction algorithms for mammography and digital breast tomosynthesis (DBT). Since these imaging techniques require the mechanical compression of the breast to obtain appropriate image quality, many such algorithms make use of breast-like phantoms. However, if phantoms do not have a realistic breast shape, this can impact the validity of such algorithms. PURPOSE: To develop a point distribution model of the breast shape obtained through principal component analysis (PCA) of structured light (SL) scans from patient compressed breasts. METHODS: SL scans were acquired at our institution during routine craniocaudal-view DBT imaging of 236 patients, creating a dataset containing DBT and SL scans with matching information. Thereafter, the SL scans were cleaned, merged, simplified, and set to a regular grid across all cases. A comparison between the initial SL scans after cleaning and the gridded SL scans was performed to determine the absolute difference between them. The scans with points in a regular grid were then used for PCA. Additionally, the correspondence between SL scans and DBT scans was assessed by comparing features such as the chest-to-nipple distance (CND), the projected breast area (PBA) and the length along the chest-wall (LCW). These features were compared using a paired t-test or the Wilcoxon signed rank sum test. Thereafter, the PCA shape prediction and SL scans were evaluated by calculating the mean absolute error to determine whether the model had adequately captured the information in the dataset. The coefficients obtained from the PCA could then parameterize a given breast shape as an offset from the sample means. We also explored correlations of the PCA breast shape model parameters with certain patient characteristics: age, glandular volume, glandular density by mass, total breast volume, compressed breast thickness, compression force, nipple location, and centre of the chest-wall. RESULTS: The median value across cases for the 90th and 99th percentiles of the interpolation error between the initial SL scans after cleaning and the gridded SL scans was 0.50 and 1.16 mm, respectively. The comparison between SL and DBT scans resulted in small, but statistically significant, mean differences of 1.6 mm, 1.6 mm, and 2.2 cm2 for the LCW, CND, and PBA, respectively. The final model achieved a median mean absolute error of 0.68 mm compared to the scanned breast shapes and a perfect correlation between the first PCA coefficient and the patient breast compressed thickness, making it possible to use it to generate new model-based breast shapes with a specific breast thickness. CONCLUSION: There is a good agreement between the breast shape coverage obtained with SL scans used to construct our model and the DBT projection images, and we could therefore create a generative model based on this data that is available for download on Github.