Development and evaluation of a retrieval system for a medical image database
The aim of this specific project is to develop and evaluate medical image retrieval based on visual content for radiologists. The retrieval finds cases (i.e., imaging volumes and corresponding clinical variables) in a curated database of medical images, based on visual information such as local appearance. It optimizes the workflow of radiologists and provides access to relevant data for individual assessment. It aims at reducing the time needed to find case specific, relevant information. Ultimately this aims at increasing in diagnosis accuracy and efficiency.
Since our first proposal in 2016, we have used NLST data with great success in our project. Granted extended usage permission, we will use the data (images & metadata) to develop and evaluate methods in two specific scenarios
1.) To develop, train and evaluate different content based retrieval approaches, built based on state of the art machine learning methods.
2.) To evaluate retrieval quality, accuracy and usability of the system. Therefore, data will be integrated into a database of reference cases from which the retrieval system selects the instances to be presented to the radiologist. A large amount of high quality images (and metadata) in the database is necessary to evaluate retrieval quality and accuracy (such as the overlap of findings between query- and result case), as well as the usability of the system. Furthermore, the data allows showing the radiologist only highly specific and relevant reference cases, thus improving accuracy and efficiency of diagnosis in clinical practice.
Rene Donner, Contextflow GmbH
Markus Holzer, Contextflow GmbH
Thomas Schlegl, Contextflow GmbH
Jakob Scheithe, Contextflow GmbH
Georg Langs, Medical University of Vienna
Allan Hanbury, Vienna University of Technology