Multi-level automatic medical image annotation

Abstract

Medical images form an essential source of information for various important processes such as diagnosis of diseases, surgical planning, medical reference, research and training. Therefore, effective and meaningful search and classification of these images are vital. Most of the medical image classification and retrieval systems use visual feature matching technique; that is extracting low-level visual features of shape, color and texture from an image and matching these features with features in the database. However, there is a semantic gap which is a gap between a low-level feature and high-level concept, the way humans interpret an image. Manual annotation is often used for medical domain image database system; that is a user enters some descriptive keywords about the image and this description is stored as metadata. However, manual annotation has problems and limitations such as domain knowledge needed by an annotator, cost incurred to annotate large amount of images, time consuming and inconsistency whereby different annotators or domain experts might use different keywords. The process by which a computer system automatically assigns keywords or concepts to an image is referred to as automatic image annotation which can provide a platform to bridge the semantic gap. Image annotation can be considered as classification problem. In addition, machine learning techniques could be used for classification. This implies that training data can be used to learn or build a classifier; and subsequently this classifier can be used to classify or annotate test images.

The main contribution in this research work is the modeling and development of framework of classifiers for multi-level automatic image annotation. The proposed framework evolves on the idea that multi-level feature extraction and concept hierarchy can improve content description of an image. In addition image retrieval is based on either text or image content. A system code-named “Medical Image Annotation and Retrieval System” (MIARS) was implemented based on this framework. The novel method of image indexing using multi-level features is also incorporated in MIARS. Experiment performance measures were conducted to evaluate the novel implementation of multilevel automatic medical image annotation framework and machine learning techniques.