A new measure for comparing biomedical regions of interest in segmentation of digital images

The segmentation of the region of interest (ROI) of digital images is generally the first step in the pattern recognition (PR) procedure. Automatic segmentation of biomedical images is desirable and comparisons among new approaches, by using available databases, are important. We present a new approach to compute the Hausdorff distance (HD) between digital images. Although HD is the most used distance estimator among sets, we show why it is not suitable for biomedical applications. In this paper, a new technique to define the degree of correction of the ROI is developed to serve as a basis for the comparisons used to validate works on segmentation of biomedical images. As for online diagnosis, the comparison among possible techniques must be efficient enough to: (1) be done in real time (i.e. during the examination), (2) allow the inclusion of priority aspects, and (3) be intuitive and simple enough to be easily followed by people with no computational or mathematical background. We develop a new index by considering the expectations of the medical doctors who are using computer systems for diagnostic aids, and take into consideration how these systems use ROIs to extract feature properties from the examinations. We discuss conditions for empirically defining a measure for calculating similarities and differences between ROIs. The proposed method is applied to both real and simulated data examples.