"Searching for the unknown" aims to find new images of the same type among large amounts of medical image data on the basis of 1 or a limited number of known examples of that type. Rather than using the advanced stochastic, physiological and geometrical models, which have revolutionized the field of medical image processing over the last two decades, we consider the analysis of these images as a machine learning problem only. As a consequence of our approach, we do not search for precise pathological facts, or the high accuracy required in medical diagnosis. Rather, we aim to search in large amount of medical image data for possible similars of the example image provided to the system. It is completely open what is searched for as long as there examples available (and visual distinctions which will hopefully be isolated by machine learning the learning phase of the system). Our motivation to research the problem is the considerable progress made in the field of semantic image classification. The categorization of random pictures into N different categories, typically "boat", "chair", "cow", but also abstract ones like "outdoor", "wedding", "head of state". The state of the art in this field is such that N may vary between N=1 to N=10,000 with average good levels of success in data sets containing upto a million of photos. What is needed to train the recognition of the concept is a number M of known examples. Apart from this coarse-grain categorization, recent progress in fine-grain categorization is capable of distinguishing 100 close resembling concepts, like bird types. With essentially the same machine learning algorithms, also the localization of the recognized objects of the categories has made considerable progress. And, finally, in instance search, large sets of data will be searched for similars starting from 1 visual example. The progress is considerable here as well. We take inspiration from these recent progresses to search in medical datasets for unknown, that is unmodelled, visual patterns on the basis of selected example patches.
We have successfully been applying state-of-the-art computer vision algorithms to medical image analysis. Preliminary results showed, that even with very little domain specific knowledge and only utilisation of generic algorithms developed for general image classification (scenes, faces, etc.), we were able to perform en par with medical image analysis techniques in Alzheimer's diseased vs. cognitive normal classification on MRI brain scans. These results are very promising and we are currently investigating ways to further improve our algorithms.