||1R21CA241705-01 Interpret this number
||Weill Medical Coll Of Cornell Univ
||Perceptual Sensitivity to Anatomical Background Statistics in Mammography
It is well known that normal anatomy in a medical image can mask the presence of disease. However
this process is not well understood. Part of the problem is that we lack knowledge of the relevant statistical
descriptors that characterize perceptual effects of image statistics. While image acquisition noise is largely
characterized by its second moments (power-spectrum or covariance matrix), background anatomy has a
complex structure that requires higher-order statistics – and an understanding of their perceptual relevance
–to characterize fully. This is an important limitation because reading “through” this background is a critical
component of many clinical tasks. In a statistical sense, reading through the background means exploiting
redundancies in the presentation of normal anatomical structures for the purpose of isolating disease
processes. The need for background characterization is well recognized in screening mammography, our
focus, as screening mammography typically includes an assessment of the background via the BIRADS
density score. However, this score has limited utility as a statistical descriptor.
The basis for this project is to translate a successful approach from basic vision science to medical
imaging, in order to identify the relevant high-order statistical properties of medical images and their
perceptual impact. In this approach, a set of local image statistics (co-occurrence probabilities) are used to
build an ”alphabet” for the statistical structure of synthetic visual textures and their local features (such as
edges). Perceptual sensitivities to local features can be concisely characterized and modeled via this
alphabet, and it has been shown that sensitivity to these elements is matched to their informativeness in
natural scenes. This motivates our general approach, and many specifics of our research plan.
Our plan is to develop algorithms in Aim 1 that selectively alter (either increase or decrease) the co-occurrence statistics of mammograms, while retaining their general background appearance. The sub-aims
explore four strategies, building on a Fourier domain approach for which a proof-of-principle is in hand.
Then, Aim 2 will use these images to assess perceptual sensitivity. Aim 2A will develop the psychophysical
paradigm. Aims 2B-D will determine whether the principles identified in previous studies of synthetic
visual textures (sign-invariance, approximate scale-invariance, and quadratic combination) extend to
medical images, as this will enable a comprehensive yet concise description of perceptual sensitivity. We will
pursue these aims using a database of full-field digital mammograms.
The project is expected to yield a validated approach for modulating high-order statistical properties
of mammograms and baseline data of perceptual sensitivity to these modulations. These findings will
improve our understanding how normal anatomy impacts the statistical properties of screening
mammograms, and give us valuable baseline data on how the statistics of normal anatomy affect perception.