The evaluation of automatic detection of structures on medical images is usually done by calculating a pixel to pixel analysis such as sensitivity and specificity analysis. However, when the sample contains a high variability, the average of these indicators is not recommended as it can hide outliers. In this article we present a weighted matching analysis which is also a pixel to pixel analysis that uses the statistical significance of observations to differentiate positive and negative pixels. It computes a probabilities map from these same observations and calculates the F-measure between the analysis and the statistics image. The proposed method was applied on the evaluation of drusens automatic quantification algorithm developed by the authors. It was used to compare the automatic generated images to a set of 22 images graded by four ophthalmologists. They were also analyzed using the binary analysis of positive and negative pixels, for comparative purposes. We concluded that the weight matching analysis improves the binary image evaluation, especially on image sets that contain analysis with significant variability, and that are also useful for dynamic samples due to its ability to be easily updated.
|Title of host publication||IFMBE Proceedings|
|Publication status||Published - 1 Jan 2009|
|Event||World Congress on Medical Physics and Biomedical Engineering - |
Duration: 1 Jan 2009 → …
|Conference||World Congress on Medical Physics and Biomedical Engineering|
|Period||1/01/09 → …|