Blind Image Restoration Enhances Digital Autoradiographic Imaging of Radiopharmaceutical Tissue Distribution.

Digital autoradiography (DAR) is a powerful tool to quantitatively determine the distribution of a radiopharmaceutical within a tissue section and is widely used in drug discovery and development. However, the low image resolution and significant background noise can result in poor correlation, even errors, between radiotracer distribution, anatomic structure, and molecular expression profiles. Differing from conventional optical systems, the point-spread function in DAR is determined by properties of radioisotope decay, phosphor, and digitizer. Calibration of an experimental point-spread function a priori is difficult, prone to error, and impractical. We have developed a content-adaptive restoration algorithm to address these problems. We model the DAR imaging process using a mixed Poisson-gaussian model and blindly restore the image by a penalized maximum-likelihood expectation-maximization algorithm (PG-PEM). PG-PEM implements a patch-based estimation algorithm with density-based spatial clustering of applications with noise to estimate noise parameters and uses L2 and Hessian Frebonius norms as regularization functions to improve performance. First, PG-PEM outperformed other restoration algorithms at the denoising task ( 0.01). Next, we implemented PG-PEM on preclinical DAR images (F-FDG, treated mouse tumor and heart; F-NaF, treated mouse femur) and clinical DAR images (bone biopsy sections from RaCl-treated castration-resistant prostate cancer patients). DAR images restored by PG-PEM of all samples achieved a significantly higher effective resolution and contrast-to-noise ratio and a lower SD of background ( 0.0001). Additionally, by comparing the registration results between the clinical DAR images and the segmented bone masks from the corresponding histologic images, we found that the radiopharmaceutical distribution was significantly improved ( 0.0001). PG-PEM is able to increase resolution and contrast while robustly accounting for DAR noise and demonstrates the capacity to be widely implemented to improve preclinical and clinical DAR imaging of radiopharmaceutical distribution.