Development of simulation software for digital mammography data to determine the image quality via a deep learning observer

Department 6.2, Dosimetry for Radiation Therapy and Diagnostic Radiology, is developing software for simulating realistic mammography data in cooperation with the Radboudumc and the LRCB in Nijmegen (NL). This software is intended to create a database of mammograms to train a deep learning observer for determining the image quality of mammography devices from different manufacturers using images with an anatomical background.

During breast‑cancer screening by means of mammography, predominantly healthy women are exposed to radiation, even if it is only a comparably low dose. For this reason, good image quality at the lowest possible dose is particularly important. The measurements for quality assurance are performed as prescribed using technical phantoms that do not bear any resemblance to a female breast [1] (see Figure 1a).

Modern mammography devices, however, have long since developed from simple linear imaging devices to complex computers. Before a radiologist examines a mammogram, internal software has already processed the original image (also called “for processing”) to such an extent that suspicious structures are more easily visible and irrelevant aspects move to the background (also called “for presentation”, see Figure 1b). What is a huge help for making a diagnosis, becomes a problem for determining the image quality. Due to the nonlinearity of the processing of the original image, the data processing step needs to be included in the determination of the image quality. However, this is not possible with the currently used technical phantoms, as the software can only process realistic breast images.

One approach to this problem is to use realistic (anatomical) phantoms for the quality measurement. This development, however, is not easy and requires some preliminary steps. Some important first measures here are the simulation of such phantoms and their corresponding realistic mammograms, and the development of software that can determine the image quality with sufficient certainty using these phantoms.

PTB’s Medical Imaging Working Group, in cooperation with the Radboudumc and the LRCB (both in Nijmegen, NL), is currently developing such simulation software for generating mammography data that corresponds to the output of devices from different manufacturers. In order to simulate a physically realistic imaging process, similar to that of a mammography device, the real processes taking place have to be reproduced as accurately as possible. Based on digital breast phantoms, the attenuation in the material passed by each examined incident ray is determined using ray tracing [2], and a primary image is generated using a realistic X‑ray spectrum [3, 4]. The effects of scatter, noise and limited detector resolution have not been taken into account yet in this primary image. These parameters are added in a system‑specific manner a next step. Both resolution and noise are adapted by means of the device‑specific MTF and the NPS, respectively [5]. A neural network that has been trained on Monte Carlo simulations is to be used for modelling the scatter. An image simulated in this manner would correspond to a real “for processing” mammography image. Another neural network is to be trained, which is to ensure the device‑specific processing of the raw images to “for presentation” images (see Figure 1c).

By using the simulation software described above, a database of many different mammograms can be created to develop a deep learning observer for determining the image quality via machine learning. This deep learning observer, apart from the already mentioned advantage of including the image processing step in the image quality measurement, is to additionally eliminate another disadvantage of the current quality measuring method. Currently, a large number of images are required to determine the image quality. A doctoral thesis that was written as part of a cooperation initiative between PTB’s Medical Imaging Working Group and its Data Analysis and Measurement Uncertainty Working Group has come to the following conclusions: The research results have shown that taking one single image would be sufficient for a deep learning observer applied to technical phantoms [6].

We hope to achieve the same results with the new method applied to anatomical phantoms.

Figure 1:

a    Image of a contrast‑detail mammography (CDMAM) test specimen for determining the image quality in mammography [6]. It consists of a grid structure and gold dots with different diameters and thicknesses. A contrast‑detail curve can be determined based on this.
b    The processing software inherent to the device converts an image “for processing” (left) into an image “for presentation” (right). This image processing is a nonlinear process and leads to the fact that no clear relationship can be established between measurements on the technical phantom and real breast images.        
c    Schematic representation of the simulation software for mammography data.

References

[1]   Perry, N. et al. (2006), European Guidelines for Quality Assurance in Breast Cancer Screening and Diagnosis, 4th ed., European Communities, ISBN 92-79-01258-4.

[2]   Siddon, R. (1985), Fast calculation of the exact radiological path for a three-dimensional CT array, Med.Phys. 12(2), doi: 10.1118/1.595715.

[3]   Hernandez, A. et al. (2017), Generation and analysis of clinically relevant breast imaging x-ray spectra, Med.Phys. 44(6), doi: 10.1002/mp.12222.

[4]   Ketelhut, S. et al. (2020), Catalogue of x-ray spectra of Mo-, Rh-, and W-anode-based x-ray tubes from 10 kV to 50 kV, Phys Med Biol. 66(11), doi: 10.1088/1361-6560/abfbb2.

[5]   Saunders, R., Samei, E. (2003), A method for modifying the image quality parameters of digital radiographic images, Med.Phys. 30(11), doi: 10.1118/1.1621870.

[6]   Kretz, T. et al. (2020), Mammography Image Quality Assurance Using Deep Learning, IEEE Trans Bio Eng., 67(12), doi: 10.1109/TBME.2020.2983539.

Contact

F. Mauter, Department 6.2, Working Group 6.24