X-ray Diagnostics Breast Tomosynthesis and Mammography
Members of the LUCI (Lund University breast Cancer Imaging) group from Medical Radiation Physics and Diagnostic Radiology
Breast tomosynthesis (BT) is a form of tomographic imaging procedure, compute tomography adapted for breast diagnostics or simply put: 3D-mammography. Tomography involves taking images from different angles and combining the information in them to create an image volume where every slice can be individually reviewed. For breast imaging this is especially important, as tumours can be hidden by overlapping breast tissue. The key difference compared to normal computed tomography (CT) is however that breast tomosynthesis only uses a limited angular range (e.g. 50 degrees) instead of a full 180-degree range. This limits the depth resolution of breast tomosynthesis, resulting in what can be called pseudo-3D. But, the big advantage of viewing imaging data slice-by-slice remains so that tissue overlap is a substantially smaller problem and more tumours can be detected.
Digital mammography is the current standard in breast imaging. Breast imaging imposes several specific requirements on x-ray equipment, e.g. being able to distinguish between soft tissue with very similar radiological density and also having a very high resolution (>100µm) to see small specs of calcium, so-called microcalcifications, and other subtle changes in breast tissue. As breast imaging is mainly a screening modality it is also vitally important to keep radiation doses low (as most women who undergo it are healthy) and that image review is fast and efficient. Breast tomosynthesis has the same requirements.
In Malmö extensive research is carried out in the field of breast tomosynthesis and mammography. In 2006 the first prototype system was installed, and in 2010 one of the world’s first commercial systems was put into use. Research is conducted in close collaboration with Diagnostic Radiology in several fields, including but not limited to screening, optimization and simulation. The joint group is known as LUCI, Lund University breast Cancer Imaging
In 2010 a large randomized study on the effect of using breast tomosynthesis in screening was started – the Malmö Breast Tomosynthesis Screening Trial. Patient inclusion was completed in 2015 with the last of planned 15000 women imaged. All included women were imaged with both mammography and breast tomosynthesis in a paired study design. The study is unique in that it investigates breast tomosynthesis as a stand-alone modality rather than in combination with standard mammography and in that it, to save dose and review time, uses only a single projection instead of the two that has been the standard for mammography.
The results of the finished study (Zackrisson et al. 2018, Lancet Oncology) showed a 34% increase in the number of cancers detected using breast tomosynthesis, evenly split between different tumour sizes, breast densities and tumour types. This is very similar to the results of other similar studies using breast tomosynthesis in two projection together with mammography in two projections. The study does however show and increase in false positives, which will be the subject of upcoming research.
Various projects aimed at optimizing breast tomosynthesis have been performed in Malmö. There is no consensus on optimal values for dose, dose per projection, angular range and many other parameters. A focus has been on analysing different viewing conditions, among other things to possibly improve the efficiency of the rather time-consuming review of breast tomosynthesis images. The separation between slices is one such parameter, where studies have investigated if a smaller number of slices can improve review time without impairing detection performance. 2019 has also seen the start a project aimed at optimizing and comparing BT systems from different manufacturers.
In collaboration with the Humanities lab in Lund we have used so-called eye-tracking in several projects, and more are planned. Eye-tracking is used to directly follow the gaze of a user on a screen. It can be a very effective way of checking what features in an x-ray image draws attention, and if there are e.g. differences between experienced and less experienced readers. One important result from our earlier research is that it is better to view images of the breast horizontally than vertically, as peripheral vision has a greater extent in the longitudinal axis and the image is thus better adapted to the field of vision.
Even if human studies are preferable, it is not always appropriate or even possible to use test subjects. This is especially applicable to the field of radiology as one cannot and should not expose large populations to ionizing radiation. Therefore, and also because of time constraints, it is often important to carry out studies on phantoms and especially digital phantoms when e.g. several different imaging parameters need to be evaluated.
From 2013-2018 an extensive project to develop a full-scale simulation of a DM/BT-imaging system was carried out. We employed the open source Monte Carlo and ray tracing software PENELOPE and PENeasy, developed by the US FDA (Food and Drug Administration). Both verification studies and simulations were carried out of different imaging geometries, trying to find and optimal trade-off between angular range (which affects depth resolution) and dose per projection (which affects contrast).
Simulated imaging requires not only simulated imaging equipment but also simulated objects to be imaged, i.e. software breast phantoms in our case. We have worked for several years in modelling breast tissue and lesions, both soft tissue tumours and clusters of micro-calcifications. These models have been used in several studies meant to define detectability thresholds for different findings by simulating imaging at different dose levels. Since 2013 we have collaborated with the University of Pennsylvania, using their breast phantom in several different studies. We have also modified the phantom to appear more realistic by filling it with breast tissue based on an in-house developed implementation of the Perlin Noise algorithm, which generates coherent and continuous structures that can be easily modified to fine tune their appearance. This method has proven to provide structures that appear similar to a range of different breast tissue types.
Virtual Clinical Trials
In a further collaboration with the University of Pennsylvania we are currently building up our knowledge in a new field, Virtual Clinical Trials. VCTs combine the disparate simulation steps described above into a package which can be used to simulate both imaging modalities, organs, lesions and readers, in essence running entire trials on the desktop instead of requiring large and expensive trials on human subjects. While not intended to replace real clinical trials, VCTs are a powerful tool to investigate effects of changes in parameters and modalities. We will develop VCTs for the combination of breast tomosynthesis and mechanical imaging.
Compression and mechanical imaging
In 2010 we conducted a study on the effect of reducing breast compression force at BT to half the standard value. It showed that effect on image quality was negligible and the difference in thickness surprisingly small. This intriguing result gave rise to new project where thin sensors are used to map the distribution of pressure on the breast surface.
Our studies have shown that the average increase in thickness incurred by reducing the average compression force from 130 N to 70 N by 1.8 mm (with an average breast being 55-60 mm thick when under compression). In many cases, the force is distributed exclusively to the innermost part of the compressed breast, mainly over the pectoral muscle, with low or non-existent pressure over the rest of the breast. Even for comparatively well-compressed breasts, most of the force is on the innermost areas close to the chest wall. This contributes little to the compression of the parts of the breast which are most important from a diagnostic view point.
Another result from our studies was the increased pressure over breast lesions compared to the background. To test the applicability of this in a diagnostic situation, we launched a study where patients with suspected breast cancer were imaged with a combination of mammography and our so-called mechanical imaging, MI, where the pressure distribution over the compressed breast was recorded and matched with the mammogram. Results from the pilot study are promising, indicating a cut-off value for cancers, where no malignant tumour had a pressure of less than 40% more than the background. These results imply that if MI could be used in screening it could provide a 30-40% reduction in false positives, i.e. healthy women called back for additional examinations. A larger study to confirm these results is currently underway in Malmö and Lund and a further study on using MI together with breast tomosynthesis is in the planning phase.
In 2019 the group began a project in collaboration with Siemens Healthcare where synthetic mammography (SM) will be evaluated for use in breast screening. SM is a 2D image created from the same data as a 3D BT image volume. It is intended to be equivalent to conventional DM. There are many reasons why this is important, even if BT is a superior screening modality to DM. One is to be able to use it to compare to prior images, another to give an overview of the breast anatomy. This will substantially save patient doses by not requiring a separate DM exposure as a compliment to BT, something which is used in some screening programmes. For these purposes, the image quality should be sufficient. It does not necessarily need to be equivalent to DM, but it must be good enough.
The project will compare DM and SM in a number reader studies. Different varieties and versions of SM will be evaluated, using different contrast settings and artefact reduction schemes.
Artificial intelligence is a buzzword in the science community. Our group is currently working on evaluating the use of AI in breast screening by performing large validation studies on the well-studied MBTST material. This is mainly aimed at answering two different questions; if deep-learning based AI can be used to safely and accurately state that a substantial portion of screening cases are healthy and do not need to be reviewed by a radiologist, and if it can be used to find additional cancers.
Anders Tingberg, PhD
anders [dot] tingberg [at] med [dot] lu [dot] se (anders[dot]tingberg[at]med[dot]lu[dot]se)
+46 40 33 11 55
Daniel Förnvik, PhD
daniel [dot] fornvik [at] med [dot] lu [dot] se (daniel[dot]fornvik[at]med[dot]lu[dot]se)
Pontus Timberg, PhD
pontus [dot] timberg [at] med [dot] lu [dot] se (pontus[dot]timberg[at]med[dot]lu[dot]se)
Anna Bjerkén, PhD student
anna [dot] bjerken [at] med [dot] lu [dot] se (anna[dot]bjerken[at]med[dot]lu[dot]se)
+46 40 33 86 56
Radiology Diagnostics, LU
Predrag Bakic, Guest professor
predrag_radomir [dot] bakic [at] med [dot] lu [dot] se (predrag_radomir[dot]bakic[at]med[dot]lu[dot]se)
+46 40 33 86 59
Magnus Dustler, PhD
magnus [dot] dustler [at] med [dot] lu [dot] se (magnus[dot]dustler[at]med[dot]lu[dot]se)
+46 40 33 86 59
Akane Ohashi. Post doc
akane [dot] ohashi [at] med [dot] lu [dot] se (akane[dot]ohashi[at]med[dot]lu[dot]se)
Hanna Tomic, PhD student
%E2%80%8B%E2%80%8B%E2%80%8B%E2%80%8B%E2%80%8B%E2%80%8B%E2%80%8Bhanna [dot] tomic [at] med [dot] lu [dot] se (hanna[dot]tomic[at]med[dot]lu[dot]se)