The statistical analysis performed in order to assess the similarity among image items is based on a hierarchical Bayesian model-based data analysis approach. The RGB signal level model p(D) is linked to a high-dimensional primitive descriptor level p(P) taking into account the color as well as geometric information extracted from the data and in turn to a secondary, lower-dimensionality, independent synthetic descriptor level p(S) by conditional probabilistic links p(P|S), p(S|P). Inference is conducted in order to obtain the posterior density p(S|P) p(P|D) p(D). The obtained densities are compared with each other taking into account a bank of distance and divergence measures to carry out an association level search procedure. Category search is employed to limit search results discriminating among retrieval outcomes of different natures. Standard measures of retrieval performance are considered in the development and evaluation of the system performance. Clusters are defined by the available supervised diagnoses attached to each of the items. The centers of mass of each of the clusters can be calculated, as the relative distances of the different centers of mass provide a further measure of the quality of the ranking output by the retrieval system. The methodological strategies used for images retrieving and archive access have been described in detail in a previous publication [31].

The FIDE system, indeed, is effective in retrieving PSL images with known pathology visually similar to the image under evaluation. This performance gives a helpful and valuable aid to the clinician in the decision-making process. Indeed, we argue that a system, able to retrieve and present cases with known pathology similar in appearance to the lesion under evaluation, may provide an intuitive and effective support to clinicians which potentially can improve their diagnostic accuracy. Nevertheless, based on its ability to retrieve images similar to the one chosen as query, FIDE can be useful as a training tool for medical students and inexpert practitioners. In fact, by using it would be possible to browse large collections of PSL images using their visual attributes. The system will allow the user to mark retrieved images as positive and negative relevance feedback. Finally, the proposed system can be used to create appropriate CBIR Web Services that can be used remotely to perform query by example in various PSL image databases around the world and can be a very good complement to text-based retrieval methods [32].