This project is an extension of previous work involving the creation, development and testing of image-processing techniques designed to improve diagnostic performance. Current work has centered on methods relevant to the processes of radiologic image subtraction and tomosynthesis. Work continues in the area of automatic manipulation to facilitate registration. Particular emphasis has been placed on: 1) methods for automatically segmenting radiographic images into regions of diagnostic interest, in anticipation of the automated detection of associated lesions, 2) methods for quantifying the apparent size of lesions from these images, 3) methods for increasing the efficiency of complex, spatial-frequency-dependent manipulations essential for optimization of diagnostic performance of specific tasks. The recognition and delineation of areas showing trabecular bone was set as a primary target because of its importance in the diagnosis and monitoring of periodontal diseases. Initial efforts have been made in an approach of combining the economics of quad-tree image characterization with the split-and-merge procedure for image segmentation. An algorithm used to detect lesion boundaries has been improved. The new approach makes use of a nonparametric test for change of distribution based on the Mann-Whitney statistic. This change makes the algorithm considerably more robust in the presence of large image noise. Another improvement developed this year permits absolute quantification of lesion volume expressed in cubic millimeters of an equivalent to homogeneous bone. The standard error associated with this process was measured in a controlled study and found to be approximately 1.3 mm-3. Future activity will continue coordinate image-processing efforts with research directed toward the development of complete diagnostic systems.
