In 2006 digitisation of the Dutch breast cancer screening has started. All screening mammograms will be stored in one national archive, which will be facilitated by the use of broadband technology. As a consequence, a large database of breast cancer cases will become available in a few years. This provides a unique opportunity for the development of decision-support in this domain. A major cause of missing breast cancer cases is interpretation failure. There is strong evidence that interpretation failure is a more common cause of missing significant lesions in screening than perceptual oversights. From audits it is known that in the Netherlands more than 25% of all cancers detected in the screened population show relatively clear signs of abnormality on previous screening mammograms, while another 25% show minimal signs.
There is evidence that computer-assisted detection (CAD) of lesions in mammograms can be of help to radiologists in interpreting whether a lesion is malignant or benign. The aim of this project is to use Bayesian networks and Bayesian classifiers to further address the problem of interpretation failures by radiologists. However, interpretation of lesions requires more medical background knowledge than is currently taken into account in CAD systems. This problem is addressed by a tight collaboration between radiologists and computer scientists.
This subproject focuses on improving feature extraction in mammograms. Detection of breast cancer in mammograms can be modelled as a multi-stage process. In a first step a search is carried out to identify locations of interest in the images. Sensitive methods for automating this step have been developed in the past and will be used in this project. These methods comprise detection of masses, microcalcifications, architectural distortion, and asymmetry, which are all signs of breast cancer. There is a need for the further development of image feature extraction and standardised data representation based on classification of local image features in single views. By combining information extracted from different views, we hope to be able to improve interpretation of mammograms.
This subproject focuses on the development of methods that allow incorporation of radiological background knowledge in constructing Bayesian networks. Background knowledge is expected to play a role both in the elucidation of the appropriate Bayesian network topology as in finding appropriate context-specific dependence information. Relational probabilistic models and similarity networks have been chosen so far as a starting point for this line of research.
In this subproject we aim to obtain insight into the nature of the task of detection of breast lesions, suspected of cancer. We have established a good working relationship with Preventicon (breast cancer screening foundation in Utrecht, the Netherlands) and are now collaborating with radiologists in findings out which features and combinations of features, when detected, may help radiologist in reducing the misinterpretation error.
For more publications go to [Publications] or to the personal web-pages of the project researchers.