Sohan Rashmi Ranjan MS, Presenter: Employee, General Electric Company

Vikram Venkatraghavan BEng,MS, Abstract Co-Author: Intern, General Electric Company

DICOM tags such as procedure names or study description are good indicators of what organs are expected in an image. However, reading room workflows and image analysis are stymied by the need of manually identifying location and extents of organs in any image. This step hinders radiologists’ productivity and automating workflows. Thus, our goal is to automatically localize organs in medical images if present or notify its absence.

Our method is based on learning of characteristic signatures for organs and structures of interest through texture analysis. We propose a novel uniform local Gabor binary pattern measure for defining organ signatures, and a content-based sub-image retrieval approach for organ identification. The algorithm has two modules: a learning module, and a classification module. The learning module identifies characteristic signatures of organs and structures from pre-segmented data. The classification module searches for specific signatures in a given image. We utilize several techniques to reduce search space for efficiency. In this paper, we focus on identification of Lungs, Liver, Heart and Kidneys in CT data. For learning, we obtained 13 CT image data with the organs segmented by an expert, and learned signatures for the organs. These signatures were used to localize organs in 60 CT volumes including both whole body and partial scans. To study robustness on partial scans, partial scans were also created through chopping full body scans.  

Sensitivity is the percentage of correct identification of presence of an organ, and specificity is the percentage of correct identification of absence of an organ. The sensitivity of localization of Lung, Liver, Heart and Kidneys were 100%, 97.82%, 93.10% and 89.13%, whereas the specificity of localization were 100%, 92.85%, 90.32%, and 78.57%, respectively.

The proposed texture measure is very effective in learning unique signatures for organs. We have proposed algorithm for efficient computation and learning of unique organ signatures. Finally, we have developed efficient algorithm for automatic localization of organs if present, or notifying it absence. The method is applicable to all modalities.

Automatic identification of organs can enhance radiologists’ productivity, e.g. labeling tumors in oncology, and imaging applications by organ-specific search, fusion, partitioning, scanning, etc.

Ranjan, S, Venkatraghavan, V, Automatic Identification and Localization of Organs in Medical Images.  Radiological Society of North America 2011 Scientific Assembly and Annual Meeting, November 26 - December 2, 2011 ,Chicago IL. http://archive.rsna.org/2011/11007583.html