Image normalization : a basic requirement for computer-based automatic diagnostic applications

Purpose: to formulate a method for normalizing computed tomographic (CT) lung image data as a preparation for computer-based automatic, or semi-automatic, diagnostic applications. Materials and Methods: histograms of greyscale distributions in comparable thoracic image slices from eight CT data-sets showed different modal values for normal, constituent tissues. In a given data-set, the usually consistent modes for muscle tissue, fatty tissue, spinal process and the descending aorta have a close correlation with the brightness increase necessary to bring an anterior 50x50 image region of visually normal parenchyma to a modal greyscale value of 35 an arbitrarily chosen normal reference value. A straight line equation relates the mode for muscle tissue in a data-set with the required percentage brightness correction. The validity of the processing was tested using the information dimension of noise-reduced pixel patterns, created when standard upper and lower greyscale thresholds are applied to 50x50 regions to confine the values closely around the normalized mode. An empirically based information dimension >=1.85 is taken as an indicator of "health." Results: the criterion information dimension is a useful index of normal lung parenchyma in normalized CT data-sets. Conclusion: image normalization is a prerequisite for computer-based diagnosis of CT pulmonary images.