Learning Image Context for Improved Computer Aided Diagnosis

Contextual information can often be exploited for improving assisted search and automatic target recognition ATR Previously we re ported on the development of hierarchical pyra mid neural network HPNN algorithms for automatically learning and integrating context for target detection demonstrating the appli cation of these algorithms to the problem of building detection in satellite imagery Spence et al Problems analogous to assisted search and ATR exist in the medical imag ing community For example radiologists will search for microcalci cations in mammograms for early detection of breast cancer Microcal ci cations are small and di cult to detect and contextual information e g clustering of calci cations location relative to anatomical struc ture can improve detection Algorithms to as sist radiologists in detecting these calci cations are being bundled in the form of Computer Aided Diagnosis CAD systems The Univer sity of Chicago is a leader in developing and testing mammographic CAD systems We have demonstrated a factor of two reduction in the false positive rate of the Chicago CAD by in corporating a HPNN component into the sys tem