Programming Research Group Technical Monograph PRG-105

Ralph Philip Highnam

DPhil thesis Trinity 1992, 219 pages, ISBN 0-902928-82-1

We investigate model-based image enhancement of mammographic images and demonstrate the dangers of ad-hoc image enhancement and analysis. A model of the mammographic process, including degrading factors and the breast is developed. The breast is considered to consist mainly of "interesting tissue" (glandular / fibrous / cancerous) and fat. Knowledge of both the compressed breast thickness and exposure are necessary for the mammography-specific algorithms developed. It is the degrading factors which are the basis for the model-based image enhancement. We study four degrading factors: scattered radiation, beam hardening, the spatially varying incident radiation intensity, and poor positioning of the automatic exposure control. The spatially varying incident radiation intensity can be measured and the mammographic images compensated by performing an x-ray exposure with no object present, this also provides some of the calibration data. Poor positioning of the automatic exposure control can be overcome by modeling the action of the control unit; this proves to be useful in amplifying scatter-removed signals to create decent images. Scatter is the key degrading factor since removing it allows simulation of a monoenergetic x-ray beam, and thus removal of the effects of beam hardening. We model scatter with a conjectured relationship between energy imparted due to scatter at a central pixel with energy imparted in a surrounding neighbourhood. Simulating a monoenergetic x-ray beam requires a choice of a photon energy, and this choice can be made with a consideration only of image quality, since radiation dose to the breast is irrelevant in a theoretical situation. Scatter removal is local high-pass filtering, and the monoenergetic simulation introduces contrast according to the chosen photon energy. Both algorithms greatly enhance mammographic images, including microcalcifications. The overall model can be verified quantitatively by inspection of the thickness of interesting tissue which must have been present to give the measured attenuation. The results of these tests were exceptionally good.

Also considered in this thesis are the effects of breast compression on the mammogram. A new technique is proposed ("differential compression mammography") which aims to aid diagnosis by using changes that are observed between two mammograms performed at different compressions. The initial results of a clinical trial are presented, and these are encouraging.

We show that if image analysis algorithms are to work robustly, account of the imaging parameters and degree of breast compression must be taken.