Abstract
X-ray mammography is the imaging modality of choice in screening to detect breast cancer in its early stages. In recent years, film-screen systems have been replaced by various digital mammography technologies as these can deliver better performance than conventional film-screen technology. However, it remains unclear how the physical performance of such systems and the choice of their operating parameters is correlated with the ability to detect early breast cancer. While clinical trials are used to address this issue, they have many associated limitations such as unethical extra exposure, time consuming data collection and completion of trials. Alternatively, a simulation framework whereby suitably realistic synthetic breast cancer pathology is inserted into normal clinical mammograms to form a large database can enable a more efficient comparison of multiple systems and study of technical parameters which influence the detection task. This thesis presents a novel computational model of breast mass appearance using fractal growth which can exhibit a range of lesion appearances. Masses generated using Random Walk (RW) and Diffusion Limited Aggregation (DLA) models were inserted into raw digital 2D mammograms using a physical model of the imaging process, thus avoiding ad hoc post-processing of the final image. The simulation framework accounted for local glandularity, polychromatic X-ray spectra, image degradation caused by the imaging system acquisition process, scatter and finally processing with manufacturer’s image processing software to produce realistic lesion attenuation and contrast. An ROC study of realism gave an average AUC and corresponding 95% CIs of 0.55 (0.51, 0.59) for DLA masses. This suggests that the DLA approach appears to produce a more realistic range of mass appearances compared to the RW approach, which achieved an AUC of 0.60 (0.56, 0.63). Both results demonstrate improvement compared to previously published ROC studies of realism of the simulated masses. The mass simulation models may be used subsequently as part of a tool to evaluate different breast imaging technologies (2D and 3D) and their performance in the detection task. Digital breast tomosynthesis (DBT) may have superior performance compared to 2D mammography in terms of cancer visibility, especially in dense breasts. Lesions grown using the DLA method, previously validated in 2D mammograms, were used to simulate breast masses into clinical DBT projection images. A pilot study was performed where radiologists feedback suggests that DLA masses can be successfully embedded in DBT projections and can produce visually authentic DBT images containing synthetic pathology. However, mass appearance whilst entirely satisfactory in 2D, does not always reliably infer satisfactory appearance in DBT.