This project is a principal component of a program directed at developing automated techniques for the recognition of clinically normal mammograms in high resolution digital mammography (DM). The goal is to reliably recognize 50% of the normal images while not misclassifying abnormal images as normal. Clinically, the approach can be considered as a second opinion strategy or as a work load reduction mechanism since the majority of images are normal. Specifically, this proposal involves two areas of complementary research: (1) A thorough evaluation of new multresolution statistical detection method used for identifying normal image regions at scales relevant to microcalcifications, and (2) A feasibility study directed at developing a statistical understanding of normal tissue regions at scales relevant to masses. The first area involves expanding the image with wavelet analysis into a sum of images each containing different levels of detail or scale. Each expansion component can be analyzed with simple parametric probability models as opposed to the analysis of the complicated raw image. A statistical test following from maximum likelihood arguments can be derived that will allow the determination of normal image regions at scales relevant to calcifications. The test is applied independently at the two relevant scales, and the results are combined. If all image regions pass the normality test the image can be declared clinically normal with respect to calcifications, and areas that deviate significantly from the model are considered as suspicious. The second component of this proposal involves the investigation of new image formation model. Initial evidence indicates that mammograms can be considered as resulting from a simple linear filtering process. That is the filtering induces the irregular image characteristics. A deconvolution approach results in a regular random field that can be analyzed with parametric methods. This field will be studied with aim of developing normal tissue detection methods at large scales and applied to images that contain masses in an analogous fashion to that of the first research area. Secondary benefits of this study include developing a parametric description of digitized mammograms and parametric analysis methods that will translate to many DM applications including images acquired from direct X-ray detection imaging.
| Fowler, Erin E E; Vachon, Celine M; Scott, Christopher G et al. (2014) Automated Percentage of Breast Density Measurements for Full-field Digital Mammography Applications. Acad Radiol 21:958-70 |
| Heine, J J; Velthuizen, R P (2000) A statistical methodology for mammographic density detection. Med Phys 27:2644-51 |
