Modelling the interpretation of digital mammography using high order statistics and deep machine learning

Abstract

Visual search is an inhomogeneous, yet efficient sampling process accomplished by the saccades and the central (foveal) vision. Areas that attract the central vision have been studied for errors in interpretation of medical images. In this study, we extend existing visual search studies to understand features of areas that receive direct visual attention and elicit a mark by the radiologist (True and False Positive decisions) from those that elicit a mark but were captured by the peripheral vision. We also investigate if there are any differences between these areas and those that are never fixated by radiologists. Extending these investigations, we further explore the possibility of modelling radiologists’ search behavior and their interpretation of mammograms using deep machine learning techniques. We demonstrated that energy profiles of foveated (FC), peripherally fixated (PC), and never fixated (NFC) areas are distinct. It was shown that FCs are selected on the basis of being most informative. Never fixated regions were found to be least informative. Evidences that energy profiles and dwell time of these areas influence radiologists’ decisions (and confidence in such decisions) were also shown. High-order features provided additional information to the radiologists, however their effect on decision (and confidence in such decision) was not significant. We also showed that deep-convolution neural network can successfully be used to model radiologists’ attentional level, decisions and confidence in their decisions. High accuracy and high agreement (between true and predicted values) in such predictions can be achieved in modelling attentional level (accuracy: 0.90, kappa: 0.82) and decisions (accuracy: 0.92, kappa: 0.86) of radiologists. Our results indicated that an ensembled model for radiologist’s search behavior and decision can successfully be built. Convolution networks failed to model missed cancers however.