The Role of Spatial Relationships in the Classification of Mammographic Micro-Calcifications

Supervisor: Professor Reyer Zwiggelaar (

Mammographic screening involves a detailed visual search of breast X-rays for signs of cancer. It is expected that computer aided diagnosis (CAD), with a sufficiently high sensitivity and specificity, will lead to an improvement in readers' performance. Micro-calcifications appear in mammographic images as groups of small bright blobs. Current CAD systems can detect more than 98% of malignant micro-calcifications, but at the same time detect non-malignant (i.e. benign) micro-calcifications and as such the specificity of these CAD methods needs to be improved [1]. The automatic classification of clusters of micro-calcifications tends to be based on features extracted from the distribution and individual micro-calcifications or other anatomical structures (e.g. linear structures such as ducts or vessels) associated with malignant micro-calcifications.

The main aim of this PhD project is to investigate a number of approaches to evaluate the role that spatial relationships play in the classification of clusters of micro-calcifications. This will cover the spatial relationships within clusters (i.e. between a group of micro-calcifications that form a cluster), but also with other anatomical structures/tissue (i.e. linear structures [2], such as vessels and ducts).

As a starting point the research will use three areas of research developed within the Department of Computer Science at Aberystwyth University in recent years:

a) Cluster topology
This aspect will be building on work by Zwiggelaar [3], which linked Betti numbers and connectivity of the individual micro-calcifications to the classification of clusters. Here we will incorporate generic topological data analysis approaches such as manifold learning [4] and link measures based on this to the classification of clusters. In addition, the visualisation of the connectiveness of the clusters will play a role in the visual assessment of the developed methods.

b) Scale space trajectory modelling
This area will extend the work by Dee et al. [5,6], which covers trajectory modelling for human traffic and results in probability maps for various paths in a scene. Instead of motion, we will be looking at the relationship between various anatomical structures and how these change with scale. It might well be that such an approach is closely related to persistent homology aspects (such as barcodes) [7], which in turn is linked to Strange and Zwiggelaar's manifold learning work [4].

c) Spatial aggregation
This work is based on qualitative reasoning methodologies for spatial reasoning [8] and will concentrate on building reasoning/classification approaches based on fundamental information from a) and b) above. In addition, we will investigate the potential of link-based clustering, or more precisely the associated metrics, for the classification of clusters [9]. Some of the initial evaluation will be based on simulated data. Real data used for the evaluation will be drawn from a number of publicly available databases, such as the MIAS and DDSM databases. In addition, data from the EPIC and Trueta databases will be used. The former contains sequential mammograms from the NHS BSP screening programme. The Trueta database contains full field digital mammograms. The project will be in close collaboration with the University of Girona and a default method for the detection of micro-calcifications will be based on the work by Oliver et al. [10]. It is expected that the developed work will use/contribute/expand the ontology work by Qi et al., especially their work on describing micro-calcifications [11,12] and related image features [13].


[1] H.D. Cheng, X.P. Cai, X.W. Chen, L.M. Hu and X.L. Lou. “Computer-aided detection and classification of microcalcifications in mammograms: a survey,” Pattern Recognition 36 (12), 2967–2991 (2003).
[2] R. Zwiggelaar, S.M. Astley, C.J. Taylor and C.R.M. Boggis, “Linear structures in mammographic images: detection and classification,” IEEE Transaction on Medical Imaging 23 (9), 1077-1086 (2004)
[3] R. Zwiggelaar, “Classification of micro-calcifications using Betti numbers at various scales,” Algebraic Topological Methods in Computer Science, Paris, France 7-11 July (2008).
[4] H. Strange and R. Zwiggelaar, “Parallel projections for manifold learning,” 9th International Conference on Machine Learning and Applications, 266-271 (2011).
[5] H.M. Dee, D.C. Hogg, and A.G. Cohn, “Scene modelling and classification using learned spatial relations,” Lecture Notes in Computer Science 5756, 295–311, 2009.
[6] H.M. Dee and D.C. Hogg, “Navigational strategies in behaviour modelling,” Artificial Intelligence 173 (2), 329-342 (2009).
[7] S. Weinberger, “What is…Persistent Homology?” Notices AMS, 36-39 (2011).
[8] C. Bailey-Kellogg and F. Zhao. “Qualitative spatial reasoning: extracting and reasoning with spatial aggregates,” AI Magazine special issue on Qualitative Reasoning 24 (4), 47-60 (2003).
[9] N. Iam-on, T. Boongoen and C. Price, “Link-based approach to cluster ensemble problems,” IEEE Transactions on Pattern Analysis and Machine Intelligence, accepted (2011).
[10] A. Oliver, A. Torrent, X. Llado, M. Tortajada, L. Tortajada, M. Sentıs, J. Freixenet and R. Zwiggelaar, “Automatic microcalcification and cluster detection for digital and digitised mammograms,” submitted (2011).
[11] D. Qi, E.R.E. Denton and R. Zwiggelaar, “Semantic analysis on medical images: a case study,” 19th International Conference on Pattern Recognition, 1260-1263 (2006).
[12] D. Qi, E.R.E. Denton, J.M.E. Leason, D. Othman and R. Zwiggelaar, “The evaluation of effects on breast cancer diagnoses using the mammographic semantic information,” Lecture Notes in Computer Science 5116, 307-314 (2008).
[13] D. Qi, E.R.E. Denton and R. Zwiggelaar, “Linking image structures with medical ontology information,” Lecture Notes in Computer Science 4046, 399-406 (2006).