Abstract
Visual perception has evolved to process natural images. Medical images, however, are formed by different physical processes than natural images and therefore may have different characteristics. To investigate this possibility, we compared the statistical characteristics of medical images and natural images by examining their overall spatial frequency content and local features. We analyzed two kinds of medical images of bone: radiographs and scintigrams. Images (51 radiographs, 20 scintigrams) were obtained from the public MedPix database and hand-curated to remove all labeling and artifacts. Images were partitioned into rectangular regions of interest (ROI), and, within each ROI, subdivided into 64 x 64-pixel patches, yielding 1918 radiograph patches and 1405 scintigram patches. For both kinds of images, the spatial power spectra, computed in the standard fashion, had a spectral slope of -3.4 over most of the frequency range, in contrast to the well-known spectral slope of natural images (-2). Scintigram spectra had a modest decrease in slope at spatial frequencies above 0.1 cycles per pixel. Local image statistics were extracted by a pipeline that included whitening the images based on their measured power spectra, binarizing at the median, and then computing correlations among pairs, triplets, and quadruplets of neighboring checks at several scales, and compared with a parallel analysis of natural images (Hermundstad et al., eLife 2014). The informativeness of each class of image statistic was assessed by its variation across image patches. We found that pairwise correlations are more informative in natural images than in these medical images. A similar finding held for triplets, with essentially no role for triplets in bone radiographs. We speculate that the reduced informativeness of triplet correlations in these medical images arises from the relative lack of T junctions. Overall, our analysis showed that the statistics of medical and natural images differ in many respects.