Systematic assessment of performance prediction techniques in medical image classification: a case study on celiac...

In the context of automated classification of medical images, many authors report a lack of available test data. Therefore techniques such as the leave-one-out cross validation or k-fold validation are used to assess how well methods will perform in practice. In case of methods based on feature subset selection, cross validation might provide bad estimations of how well the optimized technique generalizes on an independent data set. In this work, we assess how well cross validation techniques are suited to predict the outcome of a preferred setup of distinct test- and training data sets. This is accomplished by creating two distinct sets of images, used separately as training- and test-data. The experiments are conducted using a set of Local Binary Pattern based operators for feature extraction which are using histogram subset selection to improve the feature discrimination. Common problems such as the effects of over fitting data during cross validation as well as using biased image sets due to multiple images from a single patient are considered.