Medical image semantic annotation based on MIL

A method of medical image semantic annotation based on multi-instance learning is presented in this paper, as a result, the “semantic gap” problem which existence between the low-level image visual features and the high-level semantics is decreased to some extent. The lung CT images are used an example in study. A lung CT image is a bag of multi-instance learning in this method, the two-dimensional convex hull algorithms is used on the extraction of pulmonary parenchyma of the lung CT image, the gray and texture feature of the double lobes are calculated individually as the instance in the bag. The image semantic annotation is achieved by using the interactive mode which generated the positive and negative bags, and adopting multiple instance learning algorithms which combined expected maximum and diversified density. A fairly good semantic annotation ability of the method is indicated according to the experiment result.