L. Chen and H.L. Tang (UK)
Semantic analysis, knowledge base update, medical image, learning from feedback
The paper presents an intelligent mechanism for dynamically updating knowledge base, which serves for reasoning semantic meaning of medical images. In the database when an unknown image is analysed, it will be segmented into sub-images and then a visual feature detector will classify sub-regions into correspondent semantic features represented using meaningful labels. Due to complicated nature of understanding medical images and limitation of image processing techniques, these detected semantic features are usually conflicting with each other. Semantic analysis will be carried out to confirm or correct these labels with the help of Knowledge Base (KB) and a suitable reasoning mechanism. One of the key parts in the KB is the information about the similarity between different semantic features, which is summarised from testing data during the training stage. When image database is updated particularly when more images are added into the archive, unchanged KB normally is not able to conduct accurate semantic analysis. This paper introduces a dynamic mechanism that is able to update KB through learning from feedback. Trained domain experts can input some new images into the system, and check the results automatically generated from the semantic analysis. If majority of detected labels (or the generated textual annotations) are wrong, users can input textual feedback as key prompt to trigger and guide new semantic analysis. Otherwise if only minority of labels are wrong, users can correct them directly on graphical user interface. These selected new images are added into previous testing data, which is used to construct KB. The KB is updated based on the new set of testing data and will be used for later semantic analysis with better precision.
Important Links:
Go Back
