Multiresolution Markov Models for Detecting Radial Patterns of Spicules in Mammograms

In 2008 around 12,000 women in the UK (458,000 globally) died from breast cancer (cancerresearchuk.org). The National Health Service's breast screening programme has screened over 19 million women and successfully detected around 117,000 cancers (cancerscreening.nhs.uk) and a recent international study by the World Health Organisation concluded that one life will be saved out of every 500 women screened.

The growing quantity of mammograms due to an expanding screening programme, and the effort required to search for subtle, occasional signs of cancer, are adding increasing pressure on NHS radiologists. As such, computer aided detection methods are now becoming increasingly attractive. A compelling feature for computer-aided methods is that the computer reader does not suffer from fatigue and distractions and the present move from film to digital mammography (cancerscreening.nhs.uk) makes computer based methods more convenient than ever. They have recently shown a comparable detection rate to radiologists with only a modest increase in false positives, albeit when acting as a second reader.

There is now substantial interest in the development of advanced statistical image processing methods to deliver computer-based systems with improved and earlier diagnoses.

A particular open problem is the detection of spicules; these are abnormal radial patterns of curvilinear structures which can offer an early indication of cancerous abnormality (even where a cancerous mass is not evident). Unfortunately, current state-of-the-art computer-aided spicule detection algorithms cannot reliably distinguish between spicules and the variety of healthy curvilinear structures, such as stroma, milk ducts, and blood vessels. As a result, the algorithms either classify healthy tissue as spicules or visa versa. This is mainly due to the fact that previous attempts have relied too heavily on heuristic "thresholding" methods.

The proposed research will combine advanced image processing and probabilistic methods to detect spicules in mammograms. We will enhance curvilinear structures in mammograms using Markov random Field constrained wavelet shrinkage. Multiresolution, contrast tolerant curvilinear measures such as phase congruence and directional regularity will be computed from the wavelet coefficients. The marginal posterior distribution will be estimated via Markov chain Monte Carlo methods to infer the presence of curvilinear structures. This will then be used to shrink the wavelet coefficients associated with non-curvilinear structures. An orientation map will then be estimated using the curvilinear enhanced image (again using multiresolution Markov random field models). Finally, coarse-to-fine and probabilistically weighted least squares solvers will be used to perform phase portrait analysis of the orientation map and hence compute a spicule probability map.

The methods will be validated by publically available datasets. Radiologists will help assess the performance and usability of the software.

Radiologists will benefit by the proposed work in a number of different ways. The purpose of the enhancement deliverables will be to make mammograms easier to read whilst, at the same time, they will offer a controllable balance between the enhanced image and the original raw data to which the radiologists are accustomed to. This will provide additional functionality to current commercial computer aided processing software. The detection deliverable will advance the state-of-the-art of computer aided spicule detection. Involvement with clinical practitioners (the Co-I. and collaborating breast radiologist) from the outset will promote medical relevance throughout the project. As a result, potential clinical benefits will be rapidly identified. Together with the emphasis on code control and effective user interfaces, the timescale required for the benefits to begin to be realised will hence be minimised (possibly to months after successful project completion, rather than years).

Other beneficiaries of the work include commercial developers of computer aided detection in mammograms (such as, e.g. General Electric, who currently supply mammogram computer aided detection workstations to the NHS). A successful project would indicate to developers the need, possibility, and benefits of following a more rigorous probabilistic approach to understanding curvilinear structures in mammograms.

As imaging technologies advance, become more cost-effective, and become more pervasive in the UK and across the world, there is an increasing need for automatic image data analysis. Possible impacts of constrained curvilinear estimation can be found in other medical fields where computer-aided image processing is used. For example curvilinear structures are important in the analysis of blood vessels in retinal images, angiography, or brain MRA data. Here, again, the work can potentially benefit the end users (clinicians who use computer-aided systems) as well as the software developers.

Beneficiaries from wider areas include practitioners of computer-aided image analysis systems such as geological surveyors (enhancement of forest trails in geological survey aerial imagery), forensic scientists (curvilinear estimation and analysis of fingerprint images), and the defence community (detection of roads in synthetic aperture radar imagery).

Since the main aim of the work is to help radiologist distinguish malignancies from benign features, the proposed work has the potential to ultimately benefit the 2 million or so women who undergo mammogram screening each year in the UK. Reducing false negatives can potentially save or extend lives. Reducing false positives will save people undue stress and health services time and money.