Improved lesion detection and quantification in emission tomography using anatomical and physiological prior information

In SPECT and PET imaging, radiopharmaceutical concentration is often strongly correlated with anatomical structure. A Bayesian image reconstruction procedure is presented that uses this a priori knowledge to improve the detection and quantification of an unknown number of lesions. The a priori distribution employed encourages the emission tomography segmentation to stay close to the anatomical segmentation. Departures from the anatomical segmentation are detected by calculating and segmenting a deviances image: Let n/sub i/ be the estimated number of photons emitted from voxel i, /spl mu//sub ri/ the estimated mean activity of the region that contains voxel i, and l(/spl lambda//sub i/;n/sub i/) the Poisson log likelihood function for /spl lambda//sub i/, where /spl lambda//sub i/ is the mean of n/sub i/. The deviances are defined as 2(l(n/sub i/;n/sub i/)-l(/spl mu//sub ri/;n/sub i/)). Parts of the image having large deviances are candidates for becoming new regions. Hypothesis testing is performed to determine which of these candidates are justified by the projection data as being new regions. The procedure was tested by adding hot lesions to a bitmap of the Hoffman brain phantom and then simulating noisy projection data. Improvements in detection and quantification of these lesions were observed as compared to FBP and ML-EM reconstructions.<>