The long term goal of this proposal is to improve the early detection of lung cancer by improving the detectability and discrimination of low contrast nodules in digital chest radiographs. Nodule detection is improved by a Bayesian estimation algorithm which increases signal-to- noise (SNR)(and thus detectability) for low contrast nodules. SNR is increased by simultaneous contrast enhancement and noise reduction. Contrast is enhanced by compensating for scattered photons. The appearance of the contrast-enhanced image is natural to a radiologist since it is an extension of the appearance commonly provided by anti-scatter grids. Noise is reduced by including prior information regarding region smoothness through a Gibbs prior distribution which applies a penalty to the variation between neighboring pixels. While this penalty is strong for small variations (to suppress Poisson noise), it is weak for larger variations (to avoid affecting resolution for anatomical structure). The scatter reduced and noise reduced images allow better visualization and decrease the false positive nodule identification since the structured background is easier to interpret. In preliminary work with anatomical phantoms, SNR was increased by a factor of two. This is encouraging when compared to the improvement factor of 1.6 provided by an aggressive anti-scatter grid. Radiologists subjectively rated the images as superior. A preliminary ROC study indicates that the Bayesian processing both increases sensitivity and simultaneously decreases false positive rates. The utility of three types of prior information will be investigated: 1 )the Gibbs prior on the image 2)a line-site model in which region boundaries are estimated and variation is suppressed within but not across the boundaries (maintaining resolution for anatomical structures), and 3)a segmentation model in which region boundaries are assigned through Bayesian classification. The technique will be applied to images acquired both with and without anti-scatter grids. Parameters controlling scatter compensation and noise reduction will be optimized to maximize SNR for nodule detection. Detectability will be evaluated using human observer ROC studies. This represents the first scatter compensation algorithm for chest radiography which increases SNR. The improved early detection of low contrast nodules in chest radiographs will significantly improve the outcome probability for patients with early developing lung cancer.
