The ultimate goal of medical image analysis is to extract important clinical information that would improve the diagnosis and treatment of disease. Computer aided diagnosis (CAD) refers to diagnosis made by a radiologist using the results of mathematical image analysis methods, implemented on a computer, in making a decision. The goal of CAD is to: (1) reduce radiologists' false negative and false positive rates; and (2) improve overall reproducibility of image interpretation. Image gray levels in magnetic resonance imaging (MRI) depend on several tissue parameters. The information present in a sequence of MRI images of the same anatomical site (an MRI scene sequence) may be used for image analysis, e.g., tissue classification or characterization. Most of the conventional methods of MRI tissue characterization are based on explicit calculations of tissue parameters. They require acquisition of several images using specific MRI protocols. Noise propagation, through the required non-linear calculations, combines with the model inaccuracies and yields unsatisfactory results. We will investigate development, evaluation and application of a feature space method for MRI image analysis without explicit calculation of the tissue parameters. The objectives of this project can be partitioned into two parts: (1) development of mathematical algorithms; and (II) investigation of issues associated with MRI data acquisition. Outline of Part (I) will be as follows. First, additive noise will be suppressed while preserving edge and partial volume information, using a multi-dimensional edge-preserving filter which we have developed for MRI image restoration. Second, a multi-dimensional feature space representation of MRI data will be generated. In this representation, normal tissues will be clustered around pre-specified target positions, while abnormalities will be clustered somewhere else. This feature space will be generated using a linear minimum mean square error transformation of categorical data to target positions. Third, supervised, partially supervised, and unsupervised methods will be used to identify clusters and to segment MRI scene into the corresponding tissues. The resulting segmented scene can be used as an aid in making diagnosis and/or quantitative measurements. Finally, our technique will be compared to current MRI feature space methods. Part (II) will include: (1) methods of equipment non-uniformity and tissue heterogeneity correction (2) effects of MRI protocols and parameters on the; performance of our technique; and (3) effects of including non-conventional MRI data, such as diffusion and magnetic transfer images, in the scene sequence. The results of this investigation will be used to optimize MRI protocols and parameters. User-friendly computer programs will be generated to implement mathematical algorithms. These programs and the performance of the algorithms will be evaluated using computer simulations and phantom studies. Clinical feasibility studies will be performed using selected clinical cases of patients with brain tumors.
| Ghanei, A; Soltanian-Zadeh, H; Jacobs, M A et al. (2000) Boundary-based warping of brain MR images. J Magn Reson Imaging 12:417-29 |
| Jacobs, M A; Knight, R A; Windham, J P et al. (1999) Identification of cerebral ischemic lesions in rat using Eigenimage filtered magnetic resonance imaging. Brain Res 837:83-94 |
| Jacobs, M A; Windham, J P; Soltanian-Zadeh, H et al. (1999) Registration and warping of magnetic resonance images to histological sections. Med Phys 26:1568-78 |
