Textural Analysis Optimisation in MRI of the Breast

T. Beach, P. Gibbs, M. D. Pickles, L. Turnbull Centre for MR Investigations, University of Hull, Hull, United Kingdom Introduction Quantitative textural analysis is an established method of image classification in aerial and satellite photography. In recent years attempts have been made to utilise texture in MRI, particularly in the brain [1-6], but also in other organs such as breast [7] and liver [8]. One of the commonest methods of analysis, the spatial gray-level dependence matrices (SGLDM) technique proposed by Haralick et al [9], has been found to provide as much discriminatory power as other procedures [10]. Calculation of textural parameters using SGLDM relies on the initial construction of co-occurrence matrices which detail the joint probability of two pixels a specified distance d apart having co-occurring values i and j. The optimal value for d is unknown. Small values of d imply classification of fine textural features whereas coarser features require higher values of d. MR images are typically acquired with 12 – 16 bit accuracy theoretically spanning 4096 – 65536 gray-levels. In practice, to avoid the potential for over-ranging, acquired data spans approximately 10 bits or 1024 gray-levels that, without modification, results in a co-occurrence matrix containing over a million elements. Anatomical and pathological regions within the image clearly contain far fewer pixels (at least 2 orders of magnitude) leading to extremely poor counting statistics. It is usual, therefore, to reduce the number of gray-levels present in the image prior to textural analysis. The appropriate level of reduction is unclear, since improved counting statistics undoubtedly result but potentially at the expense of discriminatory power. This work attempts to explore textural analysis parameter optimisation in MRI of the breast.