| Περίληψη: | Breast cancer is the second most commonly found cancer in the world and the most common cancer to be diagnosed in women. Women diagnosed with early stage breast cancer in the United Kingdom are expected to overcome their disease by more than 10 years in a percentage of 80 %, while 65 % of women diagnosed with early stage breast cancer are expected to have an overall survival of 20 years or more [2]. Mammography screening is the main reason for the increase in survival rates, by assisting in early detection of breast cancer. More recently Digital Breast Tomosynthesis (DBT), based on digital X-ray mammography, has emerged as a new promising technology, , which is able to provide tomographic slices of the breast [3].
A limitation of the 2-Dimensional (2D) Mammography X-ray imaging acquisition technique is the superimposition of normal and pathological structures. Tissues and structures of the breast are projected onto the detector, which is in a plane perpendicular to the X-ray source, appearing overlapped. Thus, malignant lesions might be concealed by the presence of overlapped glandular tissue, leading to false negatives, while false positives may be produced by the superimposition of normal tissue.
DBT is a quasi-3-Dimensional imaging technique, in which multiple low-dose projection images of the breast are acquired by a digital detector during the rotation over a limited arc of the X-ray tube, providing reconstructed tomographic images. Acquisition of DBT projection images is similar with the acquisition of conventional digital mammography images, with the major difference being the rotation of the gantry over the stationary detector. The breast is also compressed in tomosynthesis and positioned on the breast support platform in the same ways as in standard mammography (e.g. cranio-caudal, mediolateral oblique views). Each reconstructed plane corresponds to a different depth of the compressed breast. Only structures which belong to each depth are in focus in their corresponding reconstructed plane, while structures belonging to different planes are blurred. Hence, the anatomical noise is reduced, and lesion detectability is improved, compared to 2D mammography.
There are three types of reconstructed objects that appear in tomosynthesis, tomographic planes (or slices), slabs and 2D synthetic images. Tomographic slices are a set of images depicting the content of breast at various depths. The images produced in DBT are planes parallel to the detector plane, with 0.5 to 1 mm space between them.
14
This distance is reported as slice thickness, while the actual thickness of these planes is zero.
Slabs are constructed by adding multiple tomosynthesis slices together, producing thicknesses of over 1 cm. Slabs offer radiologists a way to speed up the reviewing workflow of tomosynthesis images.
Synthetic 2D mammography images are derived from all tomosynthesis reconstructed planes. Synthetic 2D mammography attempts to provide a two-dimensional view of the whole breast like standard 2D mammography. It is obtained with no additional radiation dose since it is constructed with specific algorithms from the tomosynthesis image data set.
Even if tomosynthesis uses low dose projections the increased number of them (9-25) leads to an increase in dose compared to standard mammography. The radiation dose of a tomosynthesis view can range between equal and twice the radiation dose of a standard mammographic view, depending on the technical characteristics of the system. However, even if eventually the dose ratio of a tomosynthesis view and a mammographic view becomes 1, the radiation dose of the examination will be doubled, since current protocols include both mammographic and tomosynthesis views. A possible solution to the increased dose level would be the replacement of the standard 2D mammogram by the synthetic 2D mammogram [29, 14-16].
DBT has shown to improve detection of low contrast lesions (masses embedded in dense parenchyma) [8], while its contribution in the detection of microcalcification (MC) clusters is controversial [9-13]. The lack of lesion ground truth in case of clinical data renders quantitative image quality-based mammographic mode intercomparison a challenging task, while phantom studies can overcome this limitation [17-27], while criticised for lacking realism compared to clinical data.
Realistic background has been considered in quantitative image quality assessment studies [17, 22, 23, 25], however, the analysed MC cluster objects consisted of a small number of MC particles (up to 6), further being of ideal spherical or cylindrical shape, and characterized by a non-realistic imaging appearance such as, size, shape, and spatial distribution of the individual particles. Thus, the quantitative image quality evaluation of MC clusters composed of particles with realistic size, shape and spatial distribution in realistic parenchyma background is an open research issue.
15
In this thesis the TORMAM mammographic test object (Leeds Test Objects Ltd, North Yorkshire, UK) is used (15 mm thick). One half of the TORMAM test object consists of a uniform (homogeneous) background material and contains six groups of filaments, six groups of low-contrast circular details, and six groups of calcium-based particles (S1 224–354 μm, S2 180–283 μm, S3 150–226 μm, S4 106–177 μm), simulating MC clusters of realistic appearance with respect to particle size and shape, number, and spatial distribution of particles. The other half of the TORMAM test object consists of nonuniform (heterogeneous) material simulating the appearance of different breast parenchyma tissue density patterns (from low to high density) and contains six MC clusters of realistic appearance.
A 2D and DBT system (Hologic, Selenia Dimensions, Bedford, USA) was used to image the TORMAM phantom, with Auto-filter setting and ComboHD mode. The ComboHD acquires both DBT and 2D in one compression and enables generation of s2D image.
To quantitatively evaluate the image quality of MC cluster particles of the TORMAM, signal-difference-to-noise-ratio (SDNR), index was adapted for MC clusters, assessed across the three mammographic modes (DBT, 2D, and s2D).
The effect of increased scattering on MC_cluster_SDNR, associated with increased phantom thickness (mimicking compressed breast thickness), was analysed in both uniform and nonuniform background parts of the TORMAM phantom (TORMAM test object + PMMA slabs), by adding polymethyl methacrylate (PMMA) slabs to achieve phantom thickness of 20, 30, 40, 50, and 60 mm.
The effect of MC cluster size on MC_cluster_SDNR was analysed only in the uniform background part of the phantom, by considering only the fully visible clusters, i.e., the 4 largest MC clusters of the TORMAM phantom.
The effect of increasing parenchymal density on MC_cluster_SDNR was analysed only in the non-uniform background part of the TORMAM phantom. In order to support this task the surrounding parenchymal density of MC clusters was performed by histogram analysis.
Finally, analysis of both uniform and non-uniform parts of the TORMAM phantom was repeated for an “upside-down” TORMAM test object setup to consider the influence of location of the MC clusters along the z-axis (distance from detector cover) on MC cluster SDNR.
16
Signal difference to noise ratio, requires for its calculation the definition of particle target regions and particle background surrounding regions. Dealing with MC clusters composed of many particles requires accurate and repeatable automated or semi-automated segmentation methods. An algorithm capable of capturing the varying shape and size of individual particles, based on two deformable models, a level set and an active contour one [33], is used to generate particle region ground truth.
Particle ground truth is generated for each MC cluster per mammographic mode for 20 mm phantom thickness acquisition, which is considered as a reference (ideal) acquisition condition. In case of DBT reconstructed slices, MC particle segmentation was conducted on the focal plane, which was adjusted to the specific phantom part (i.e., uniform or nonuniform). In order to enable the intercomparison of mammographic modes in terms of cluster image quality, care was taken to segment corresponding/homologous particles per cluster across modes.
Results regarding the effect of increasing scatter conditions per mammographic mode, resulted in non-statistically significant differences for both uniform and non-uniform background parts. MC cluster SDNR constancy with respect to phantom thickness increase is attributed to Automatic Exposure Control system of the specific mammographic unit.
Regarding the results of MC cluster size on MC cluster SDNR in the uniform background part of the TORMAM phantom, inter comparison among mammographic modes, demonstrated DBT and 2D superior to s2D, in case of the smallest visible cluster for all phantom thickness values studied (30–60 mm). For the largest cluster size, all modes were equivocal (not demonstrating any statistically significant differences) for all phantom thickness values studied (30–60 mm). Considering the two intermediate cluster sizes, the role of 2D is enhanced with decreasing cluster size and increasing phantom thickness.
Regarding the results of MC cluster SDNR for increasing surrounding parenchymal density in the nonuniform background part of the TORMAM phantom, the role of DBT mode was highlighted as compared to 2D and s2D, in case of increased MC cluster surrounding parenchyma density and increased scattering (increased phantom thickness). Finally, comparison in terms of MC cluster SDNR between 2D and s2D mammography did not reveal any statistically significant difference for all density
17
patterns (D1, D2, D3, D4, and D5) and for all the thickness values considered (30, 40, 50, and 60 mm).
Finally repeating the above experiments for size and density effects in presence of increasing scatter conditions for the “upside-down” TORMAM test object setup , no statistically significant differences were demonstrated in case of the uniform part of the phantom confirming the z-location of the particles in the middle of the test object, as expected. Regarding the non-uniform part, no statistically significant differences were also demonstrated between the two setups of the TORMAM test object, although particles are located 8mm further from the detector cover (they are located on the bottom side of the TORMAM test object).
In conclusion, results of this thesis suggest that at present, in case of uniform background, all three mammographic modes perform similarly for large to intermediate size clusters, but cannot replace 2D in case of small size clusters and increased phantom thickness (increased scatter conditions), while more interestingly DBT seems superior in case of increased breast density of heterogeneous backgrounds, given the size of the clusters conforms to its spatial resolution.
A number of issues remains open for in this thesis to complete quantitative image quality analysis of MC clusters in the three mammographic modes, such as MC cluster shape analysis [23] are envisioned as immediate next steps.
|
|---|
