Numerical observer study of MAP-EM regularization methods with anatomical priors for lesion detection in 67Ga images

The goal of this work is to investigate to what extent the introduction of anatomical information in the MAP-EM algorithm can improve tumor detection in ⁶⁷Ga images of the chest. The anatomical data used here are simulated CT slices of the chest. The edges of the organs and/or lesions can be easily delineated from these slices. Phantom images are created using the SIMIND Monte Carlo simulation software and the mathematical cardiac-torso phantom. Images are reconstructed using MAP-EM algorithm and the detection tasks are performed using numerical observers. Four strategies were investigated. Strategy S1: no prior, the reconstructed images are only filtered using a 3D Gaussian filter post rescaled block iterative reconstruction; Strategy S2: anatomical prior employed where lesions are present in the projection data and in the anatomical data; Strategy S3: anatomical prior employed where lesions are present in the projection data but not in the anatomical data; Strategy S4: anatomical prior employed where lesions are not present in the projection data but only in the anatomical data. Two contrasts lesion/background were used, 12/1 and 22.5/1. Three values for beta, the weight of the prior term, were under investigation: 0.005,0.02,0.04. We find that for the prior weights we have investigated (beta=0.005, 0.02 and 0.04), and for contrasts lesion/background of 12/1 and 22.5/1, strategy S2 yields better results than S1, especially at the low contrast. S3 yields results similar to S1. Strategy S4 yields results similar to S2. Thus it appears that use of the edge information for possible lesions seen in anatomical data may not be a problem in terms of creating false lesions in the emission slices.