Improving medical image perception by hierarchical clustering based segmentation

It has been well documented that radiologists' performance is not perfect: they make both false positive and false negative decisions. For example, approximately thirty percent of early lung cancer is missed on chest radiographs when the evidence is clearly visible in retrospect. Currently computer-aided detection (CAD) uses software, designed to reduce errors by drawing radiologists' attention to possible abnormalities by placing prompts on images. Alberdi et al examined the effects of CAD prompts on performance, comparing the negative effect of no prompt on a cancer case with prompts on a normal case. They showed that no prompt on a cancer case can have a detrimental effect on reader sensitivity and that the reader performs worse than if the reader was not using CAD. This became particularly apparent when difficult cases were being read. They suggested that the readers were using CAD as a decision making tool instead of a prompting aid. They conclude that "incorrect CAD can have a detrimental effect on human decisions". The goal of this paper is to explore the possibility of using hierarchical clustering based segmentation (HSC), as a perceptual aid, to improve the performance of the reader.