Breast Density and Pattern: Imaging Biomarkers for Breast Cancer

Mammography is widely used for screening the asymptomatic population for early signs of cancer, and the advent of digital imaging has opened the door to the development of automated techniques, both for detecting cancer and identifying women at increased risk. We are focussing on two measures: mammographic density, which describes the quantity of radiodense and fatty tissues in a woman’s breasts; and mammographic texture, which describes the organisation and distribution of the tissues. Both are related to risk of developing cancer, and early results indicate that both will also be related to the efficacy of mammography as a screening tool.

Lay summary: We can use our knowledge of the mammography imaging process and mathematics to measure the amount and pattern of dense tissue in the breast. We are developing methods that can predict the development of cancer, tell us accurately whether treatment to reduce risk is working, and give us an indication of the most appropriate methods of imaging an individual woman’s breasts. 

Digital Breast Tomosynthesis

Digital Breast Tomosynthesis (DBT) is a new X-Ray imaging modality which provides depth information at high resolution and low dose. We have investigated the use of DBT as part of the Tommy Trial, evaluated microcalcification CAD for DBT and begun to investigate the way in which images are interpreted. We are currently developing algorithms to automatically measure tumours and breast density from DBT images.

Lay summary: One of the new breast imaging methods that will almost certainly play a part in a personalised screening programme is DBT. This gives (almost) 3D images at a relatively low x-ray dose (so it’s safe). The images are challenging when it comes to extracting quantitative information because they are made up of a number of very low dose (noisy) components, and we are looking at mathematical ways around this, but for detection and diagnosis they are very impressive [animated slide showing a direct comparison between a DBT image and mammogram of the same breast – more abnormalities appear in the DBT image than in the mammogram].

Computer Aided Detection (CAD)

Computer based methods can be used to detect potential abnormalities and present the locations as prompts to attract radiologists’ attention. In order for CAD to be successful both the sensitivity and specificity of the prompting algorithms must be high. The situation is complex, particularly when multiple algorithms are used, and this remains an active area of interest. We have conducted experiments based on synthetic images to further our understanding of the situation, and evaluated four commercially available prompting systems and conducted both retrospective and prospective clinical trials to compare single reading with CAD and the current standard clinical practice of double reading. Our current work includes developing a computer game to harness the power of citizen science to inform future CAD systems.

Lay summary: Computer aided detection (CAD) is a method where computers find areas of mammograms that look abnormal and mark them with prompts. The radiologists and radiographers that read the images can look more closely at prompts to see if they are marking genuine abnormalities. We have shown that this can improve cancer detection, but there’s a problem. Because most mammograms don’t show cancer, and because the computer prompts things that are not cancer as well as things that are, readers end up looking at several tens of ‘false’ prompts before they find one marking a cancer, and this can lead to women without cancer being recalled from screening. It’s possible to alter the performance of the computer to reduce the number of false prompts, but this also reduces the proportion of cancers that are prompted. We need to find the best balance for prompting systems, and are doing this using a computer game where people have to detect bats in images of flocks of birds. Some players will be prompted and some won’t, and the prompt performance will be different for different players, and different levels of the game. This is being launched on the Prevent website later in the summer.

Automatic detection of mammographic abnormalities.
We have investigated the use of Bayesian statistics to improve specificity in the detection of microcalcifications, which are one of the earliest signs of cancer, and found that combining different cues was effective in improving detection (Astley & Taylor 90). The detection of asymmetry is is more difficult, as the breasts are variable in appearance and differ naturally. The technique we have developed is based on identifying anatomically similar regions using the transportation algorithm to measure similarity (Miller and Astley 93, Board et al 04).  We have also investigated the detection of spiculated masses and distortion (Zwiggelaar et al 04). One of the key ideas behind this approach was model-based classification of linear structures in the digitised images (Parr et al 96).

Lesion Modeling

We have developed a method for generating realistic synthetic masses by statistically modelling spiculated breast lesions. Many of the synthetic lesions are indistinguishable from real lesions by consultant breast radiologists (Caulkin and Astley 00, Berks et al 08a). We have also looked at the relationship of the lesions to normal breast tissue (Berks et al 08b).

DME Lead for Intercalated Degrees

DME Associate Lead for Personal Excellence Pathway (Teaching and Learning)

Prof Gareth Evans

Prof Tony Howell

Dr Sacha Howell

Dr Anthony Maxwell

Prof Ryan Halter (Dartmouth College, USA)