Texture Synthesis using Hidden Markov Measure Fields

M Rogers and C J Taylor, "Texture Synthesis using Hidden Markov Measure Fields", Meeting of the Medical Images and Signals to Clinical Information Inter-disciplinary Research Consortium, Oxford, 2005. Non-reviewed submission. PDF


Statistical models of shape and appearance are well-known to provide a powerful constraints in many areas of medical image analysis. However, current modelling techniques still have some important shortcomings that limit their applicability to some medical image analysis problems. Specifically, no adequate representation of image features lying somewhere between consistent structure and fine textural detail has yet been developed. The aim of our work is to develop models capable of representing this strextural data. Advances in this area would benefit medical applications such as breast cancer screening using digital mammography, diabetic retinopathy and CT chest scans for lung cancer screening, all of which produce images that contain highly detailed textures with complex structure. In this preliminary work, we have investigated methods of generating ergodic texture. Our aim is to continue the work of Rose and Taylor [1] on representing mammographic textures. Their method is based on an MRF formulation of texture. We have applied the Hidden Markov Measure Field (HMMF) method of Marroquin [2] to texture synthesis. The HMMF method sets the problem of MRF segmentation within a formal optimisation framework. We are able to use the method to avoid the slightly ad-hoc pixel or region-wise texture generation scheme used by Rose and others. By using HMMFs, we also hope to avoid some of the other problems associated with MRF texture synthesis process, such as texture boundary inconsistencies and invalid texture growing.