Perception and Inter-Observer Variability in Mammography

点击翻译按钮获取中文摘要

DESCRIPTION (provided by applicant): Inter-observer variability in mammogram reading has been well documented in the literature. Various factors have been used to explain this variability; among them, the most significant are related to the management of perceived findings. However, the nature of this inter-observer variability has not been explored. Namely, were the lesions that were consistently reported by the radiologists any different from the ones that yield disagreement? Furthermore, could these differences be quantitatively assessed? Moreover, were these differences in any way related with the experience level of the observer? In addition, the interpretation of perceived findings is closely related with the visual search strategy used to scan the breast tissue, because observers compare perceived findings with the background, in order to determine their uniqueness. Hence, what is the effect of visual search strategy on inter-observer variability? Can this effect be modeled using Artificial Neural Networks (ANNs)? Can inferences be made regarding the observers' decision patterns by analyzing the results of simulations run on the ANNs? The work described here aims at answering these questions. We will use spatial frequency analysis to characterize the areas on mammogram cases where mammographers, chest radiologists with experience reading mammograms and radiology residents at the end of their mammography rotation, indicate the presence of a finding, or fail to do so. We will assess inter-observer agreement, as well as intra- and inter-group agreement for the various groups of observers. In addition, we will train artificial neural networks to represent each observer, in such a way that by changing the nature of the features input to the ANNs we will be able to simulate how such changes would have affected the actual observer. We will assess the effects on inter-observer variability of changing the search strategy used by the observer to sample the breast tissue. In our setting, the inter-observer variability will be assessed by comparing the outputs of the ANNs that represent each observer. In addition, the changes in sampling strategy will correspond to actual possible strategies for the human observers themselves.