The degree of symmetry between left and rightbreast X- rays is an important indicator of the presence of breastcancer, but it is also one which is difficult to assess, both by visualinspection and automatically. This project will firstly investigate therange of normal and abnormal asymmetries, and then seek to develop anew, robust method for automatic detection. Previously we have used amulti-resolution transshipment algorithm for detection, but it isapparent that transshipment of densities alone is not sufficientlysensitive to detect all forms of abnormal asymmetry. The project willbuild on previous work, and also investigate the segmentation of breastparenchymal patterns as a basis for comparing left and right breastimages.

Benign or malignant? Analysis of detected abnormalities in breast X-rays
(Supervisor: Dr S Astley)

Computer-aided detection (CAD) systems automatically detect potentialabnormalities in mammograms, and present the locations detected to thehuman film reader as prompts. The reader must decide whether or not toact on prompts, as the specificity of CAD is currently poor and manynormal regions are marked. This project will investigate methods fordeciding whether or not marked regions correspond to genuineabnormalities with a view to (a) improving the specificity of CADalgorithms by post-processing and (b) providing diagnostic informationfor the reader as decision support.

Computer-aided training for mammography
(Supervisor: Dr S Astley)

Radiologists need to read over 5,000 mammograms per year in order tomaintain a sufficiently high level of film-reading performance. Inlarger screening centres this is feasible, but in smaller practices andat an early stage of training there is a need for an effectivecomputer-based training aid. This project will develop an adaptivetraining system that will recognise a trainee's individual strengthsand weaknesses from their performance and hence provide appropriatetraining material. It will use a database of real abnormalities andexample cases, and will also incorporate synthetic tumours built fromstatistical models.