Toyohiko Sakai MD, Presenter: Nothing to Disclose

Kenji Kondo, Abstract Co-Author: Employee, Panasonic Corporation

Kazutoyo Takata, Abstract Co-Author: Employee, Panasonic Corporation

Kazuki Kozuka, Abstract Co-Author: Employee, Panasonic Corporation

Masakai Kiyono, Abstract Co-Author: Employee, Panasonic Corporation

Hirohiko Kimura MD, PhD, Abstract Co-Author: Nothing to Disclose

Masato Tanaka PhD, Abstract Co-Author: Nothing to Disclose

The use of diagnosis support systems, which show similar cases from medical archives, has been proposed in recent years. However, such systems can mislead radiologists if similar cases are not appropriately selected. In this presentation, we have proposed a novel image-retrieval system that employs a weighting technique based on radiologists’ knowledge to select reference images and demonstrated an actual system that employs this principle.  

To compile the data required for weighting, radiologists marked 1026 regions of interest (ROIs) on lung CT images and classified them into 12 imaging patterns, including consolidation and wide-spread ground glass. Image similarity was primarily calculated using 413 types of image features. In the new technique, the calculation was weighted on the basis of the regression coefficients for each lesion pattern classified by radiologists. Similar images were retrieved with or without a weighting technique. Finally, image similarity was graded on a 5-point scale (with a score of 5 indicating "very similar") by 2 radiologists in a blinded manner. For each evaluation, 60 ROIs that included the 12 patterns were extracted as a query. The precision (subjective evaluation score ≥ 4) was 78.0% and 75.0% using the proposed method with a weighting technique, and 62.7% and 62.0% without it. The 2 sets of values showed statistically significant differences (P < 0.01).  

Lung disease generally involves diverse lesion patterns, and diverse image features are used in queries. Therefore, a simple combination of all image features results in degradation of search performance. Subjective evaluation clearly showed that the use of a weighting technique can solve this problem.

A novel image-retrieval technology is introduced in this study. This diagnosis support system, which uses weighted image retrieval with an archive of images, might help radiologists make rapid and appropriate diagnoses and consequently encourage them to make greater use of these systems.

Sakai, T, Kondo, K, Takata, K, Kozuka, K, Kiyono, M, Kimura, H, Tanaka, M, Demonstration of a Novel Diagnosis Support System for Lung Lesions: Computed Tomography with an Image-retrieval Technology Based on Radiologists' Knowledge.  Radiological Society of North America 2013 Scientific Assembly and Annual Meeting, December 1 - December 6, 2013 ,Chicago IL. http://archive.rsna.org/2013/13014299.html