Image quality in breast tomosynthesis and full field digital mammography

• Κύριος συγγραφέας: Μπάρα, Σταυρούλα
• Άλλοι συγγραφείς: Κωσταρίδου, Ελένη
• Μορφή: Thesis
• Γλώσσα: English
• Έκδοση: 2019
• Θέματα: Digital mammography; Digital breast tomosynthesis; Quantitative image quality analysis; Ψηφιακή μαστογραφία; Ψηφιακή τομοσύνθεση μαστού; Ποσοτική αξιολόγηση ποιότητας εικόνας; 618.190 757 2

Mammography is the most commonly used technique for the detection and early diagnosis of breast cancer. Through the decades, mammography has been evolved from Screen- film Mammography to Full Field Digital Mammography (FFDM) and currently to Digital Breast Tomosynthesis (DBT). DBT provides quasi- three-dimensional information via reconstructed image slices produced by a series of projection view images acquired during the x-ray source movement along a limited arc trajectory. The main advantage of DBT is that it has the potential to overcome the major limitation of FFDM, which is tissue superimposition, since it produces slices of the breast at various depths. A major issue that has emerged is the way that DBT should be used in clinical routine; as a stand- alone technique replacing FFDM or as a complementary technique to FFDM. Therefore, comparing the image quality between the two modalities is of interest and the subject of various studies reported in the literature, both clinical and phantom based. This study focuses on the quantitative evaluation between DBT and 2D FFDM mammography, using the Signal Difference to Noise Ratio (SDNR) as an image quality index and mammographic phantom images. It also aims to the investigation of the contribution of different of SDNR calculation approaches suggested in the literature, when analyzing phantom object images simulating breast lesions. The DBT and 2D FFDM images were acquired using the TOR MAM mammographic phantom (Leeds Test Objects Ltd, North Yorkshire, United Kingdom). TOR MAM consists of two parts, one of heterogeneous and one of homogeneous background. The heterogeneous background part simulates the appearance of actual breast parenchyma, containing microcalcifications in addition to fibrous and nodular details. The homogeneous part contains details that simulate pathological breast features, like masses of different nominal contrast values (4%, 3%, 2%, 1.5%, 1% and 0.5%) and microcalcification clusters of varying particle sizes (upper and lower particle size limit of each cluster: 354-224 μm, 283-180 μm, 226-150 μm, 177-106 μm, 141-90 μm and 106-63 μm). Three phantom thicknesses were considered, namely 25mm, 45mm and 65mm, simulating a thin, medium and a thick compressed breast thickness, respectively. The system used to acquire FFDM and DBT scans of TOR MAM phantom is a commercially available mammography system (Hologic, Selenia Dimensions, Bedford, USA), located at University Hospital of Patras. The acquisition exposure conditions (tube voltage and load) were selected automatically by the system. In the 2D mode, the system also selected the filter used (Rh or Ag,) depending on the compressed breast thickness. For the compressed thicknesses used in this study, all acquisitions were performed using Rh filter. In DBT mode, the system used Al filter for all acquisitions, since it was the only available filter material. The analysis focuses on the 5 highest nominal contrast values of the mass-like objects, as well as on the 4 biggest particle sized microcalcification clusters, located in the homogeneous part of the TOR MAM phantom. The SDNR index was calculated for these lesions on both 2D and focal plane DBT reconstructed images. Three (3) different approaches for mass SDNR calculation and two (2) approaches for microcalcification cluster SDNR calculation were implemented in the ImageJ software environment, that assume different definitions of target and background regions. The results of the comparison between DBT and 2D mode, according to all mass SDNR methods utilized, suggest that DBT performance is statistically significantly increased compared to 2D, for all nominal object contrasts studied and all PMMA thicknesses. Additionally, as the nominal object contrast decreases or as the phantom PMMA thickness increases, mean SDNR values are decreasing, as expected. Specifically: (a) for Method 1 SDNR values provide a percentage gain for DBT in the range of 89.5%- 50% for simulated breast thicknesses 25mm – 65mm, (b) for Method 2 the percentage gain for DBT is in the range of 77%- 47.3% for simulated breast thicknesses 25mm – 45mm and (c) for Method 3 the percentage gain for DBT in the range of 57%- 52.5% for simulated breast thicknesses 45mm– 25mm. As regards analysis of the effect of different mass SDNR methods, results indicate that in 2D mammography all 3 methods seem equivalent. However, in DBT, method performance is differentiated, with mass SDNR method 1 resulting in statistically higher SDNR values, which are reduced with method 2 and become further reduced with method 3. In case of microcalcification clusters, SDNR results demonstrate different trends depending on the method used, with respect to particle size and PMMA thickness. According to Method 1, DBT performs statistically significantly better than 2D only in case of the largest microcalcification cluster A (upper and lower particle size limit: 354-224 μm), for all compressed breast thicknesses studied, providing percentage gains in the range of 51.5% and 19.5%, while for cluster D (177-106 μm) for 25mm thickness and cluster C (226-150 μm) for 45mm and 65mm thickness, a statistically significant difference is observed in favor of 2D, with DBT providing a loss in the range of -26.6% and -17%. According to Method 2, DBT and 2D perform equivalently, except for cluster D for 25mm thickness and cluster C for 65mm thickness, where statistically significant difference are observed in favor of 2D, with DBT presenting a percentage loss of -23.2% and -23.8% respectively. As regards analysis of the 2 different cluster SDNR methods utilized, results indicate that cluster SDNR Method 1 results in higher SDNR values for both 2D and DBT modes compared to method 2 and is capable of differentiating the performance of DBT from 2D for the largest cluster sizes at all thicknesses, while for the smaller cluster sizes and for all thicknesses both Methods perform equivalently in favor of 2D.