Breast Tomosynthesis Simulator for Virtual Clinical Trials
Summary
The use of computer models and simulation in predicting the clinical performance of x-ray imaging devices in silico and generating synthetic patient images for training and testing of machine learning algorithms is widely accepted today.
In this study, the researchers presented a detailed description of the computational models from the open-source GPU-accelerated Monte Carlo x-ray imaging simulation code MC-GPU. Originally, the code was developed to simulate radiography and computed tomography. But now, it has been available for use as a digital mammography and digital breast tomosynthesis (DBT) device.
Recently, the code was used to image 3000 virtual breast models with the aim of reproducing in silicon clinical trial used in support approval of DBT as a replacement of mammography for breast cancer screening. The updated code performed better in various aspects such as source model, acquisition trajectory, anti-scatter grid, etc. Thus, the researchers believe that MC-GPU can simulate x-ray projections that include many of the sources of variability found in clinical images, and that the results are better in terms of input parameters.
Breast Cancer and Mammography
Early detection of breast cancer is an important factor in successful treatment. Because of this regular screening in women (even without symptoms) using mammography has been recommended for many years. Mammography is basically the process of getting an x-ray picture of one’s breasts. It’s used to detect early signs of breast cancer. Some breast cancers can be detected with a mammogram three years before symptoms occur.
Despite being such an excellent machine, mammogram also has their own limitations. For starters, it’s only a two-dimensional imaging modality so it doesn’t have the ability to completely differentiate lesions that overlap with tissues.
Other x-ray imaging machines that offer three-dimensional images are the digital breast tomosynthesis (DBT) and the breast (CT). These have been developed with the intent to improve breast cancer diagnosis. It was only in 2016 when the FDA first approved a DBT system as a replacement for mammography. Throughout the decade, advances in computational modeling have enabled the use of computer simulations for the research and development of medical imaging devices. However, despite these advanced simulation machines, computational modeling has not yet been used in the regulatory approval of DBT.
One study focused on the impact of modeling in the regulatory evaluation of future imaging devices. They aimed to reproduce the clinical trial used in their approval of the Siemens Mammomat Inspiration DBT system.
The MC x-ray transport simulation code MC-GPU
This article describes the MC x-ray transport simulation code which was developed to replicate the mammography and DBT devices in the VICTRE project. It explains the main assumptions and simplifications included in the code. It also presents the software verification and validation activities performed in order to evaluate the software performance.
What is the Monte Carlo method?
The MC x-ray transport simulation code MC_GPU was designed to replicate the Siemens device for mammography and DBT. The Software uses the x-ray matter interaction physics and material-specific interaction cross-sections from PENELOPE 2006.
Source Model
The reproduced clinical x-ray source has a focal spot with a nominal size of 300 µm. To model this extended source, the emission point of the x-rays is randomly sampled at a distance from
The user-defined source position following a 3D Gaussian probability distribution with 300 µm full-width-at-half-maximum, cropped at 2 standard deviations by rejection sampling to preserve sharpness.
The shape of the original focal spot is not exactly Gaussian, and Marshall and Bosmans reported that the 2D projection of the spot is approximately square. Nevertheless, the researchers considered that the model provides a good first approximation to the complex 3D shape of the focal spot.
Detector model
The detector model in MC-GPU has been extended to reproduce a 200-µm-thick amorphous Selenium (Se) direct-conversion detector with 3000 × 1500 85 µm pixels (the real detector has 3584 × 2816 pixels, but a smaller image size was sufficient to cover the largest breast phantoms in our studies).
The detector model includes a 1-mm-thick protective cover and a 1D-focused anti-scatter grid analytical model. The grid frequency is 31 line pairs/cm with a grid ratio of 5. Not knowing the exact composition of the grid, 65 µm lead strips and polystyrene interspace material were assumed. A binary random sampling was used to determine if the x-ray is absorbed or transmitted through the cover and the grid.
Patient and x-ray tracking models
MC-GPU uses a voxelized geometry model to represent the patient anatomy. The transport of the x-rays across the voxelized volume is accomplished using a delta scattering tracking algorithm.
Number of x-rays per image
Corresponding air kerma per history values for each energy spectra were simulated using the MC code with a geometry composed of only air voxels. As a basic model for an ideal ionization chamber, the air kerma measurement was assumed to be equal to the average energy deposited in a 2 × 2 × 2 mm3 air volume at 1 m from the source. As a result, the expected mean glandular dose in the DBT acquisitions was of the order of 1.5 mGy. The total number of x-rays for a DBT acquisition was equally distributed amongst the 25 projections.
Model assumptions and simplifications
The most relevant simplifications implemented in the software are the following:
- No secondary ionizing electron transport after photoelectric or Compton interactions inside the imaged object or the detector
- No blur introduced by electron-hole transport inside the Se detector: no charge sharing between pixels.
- Constant electron–hole-pair generation gain, W+, of 50 eV
- No fluorescence inside the patient
- X-ray energy spectrum sampled from a probability distribution function with identical spectrum emitted at every angle
- Ideal beam collimators
- Simplified interaction modeling inside the detector and anti-scatter grid
The following assumptions were implemented in the models for lack of detailed knowledge:
- No patient motion during image acquisition.
- 3D-Gaussian (symmetrical) extended focal spot with no offfocus radiation.
- No semiconductor inner structure with no blocking layer at the top and bottom sides of the detector and 100% effective fill factor
- No temporal effects on the detector response such as lag or ghosting
- No structured or fixed pattern noise in the images
- Approximate thickness and composition of the anti-scatter grid interspace (polystyrene) and strips (lead).
- Simple additive electronic noise without a detailed model of the pixel charge read-out electronics.
- No image post-processing (generating only raw images ‘‘for processing’’ not ‘‘for presentation’’).
- The number of x-rays for each breast glandular was estimated using AEC output measured with uniform phantoms with equivalent thickness, neglecting the correlation between thickness and granularity.
- Same source motion blur (i.e., same exposure time) assumed for all exposure settings.
Conclusion
The computational models used in the MC-GPU software to replicate clinical mammography and DBT imagine systems in the VICTRE trial are recorded and assessed for credibility. Verification studies also indicate that the basic computational models created to reproduce the device operation were correctly implemented in the source code.
However, differences with the clinical system remain in terms of bench testing performance, and the software cannot yet be considered fully validated for any clinical or regulatory
context of use. Methodologies that can be used to improve the computational models and to further validate the code performance were presented. The simulation software has been
released as open-source software and can be improved and extended by other groups to replicate other breast imaging devices or to study other x-ray imaging applications.
Oncology Related Tools
Other
Latest Research
