Michal Sharon MD, Presenter: Nothing to Disclose

Uri Avni, Abstract Co-Author: Nothing to Disclose

Jacob Goldberger, Abstract Co-Author: Nothing to Disclose

Eli Konen MD, Abstract Co-Author: Advisory Board, Valtech SA Founder, RadLogics

Hayit Greenspan PhD, Abstract Co-Author: Nothing to Disclose

The BoW model-based software may have an important prioritization role in the future workflow of imaging departments. Further refinement of this technology is required.

The increasing use of diagnostic imaging and shortage of radiologists may result in long reporting time delay in busy medical centers. A computerized automated system which could analyze chest films and draw early attention to abnormal findings would be desirable. Our purpose was to present a software which automatically identifies abnormal chest films and to present our preliminary clinical experience in a busy ER. The software is based on the ”Bag of Words” (BoW) model: the chest film is divided into numerous small patches. Patches taken from a training set of chest films are clustered automatically to form a “dictionary of visual words”. For a given new chest film, each pixel patch is transformed to the index of the closest word. Each film is thus represented as a unique histogram on the dictionary word set. Multiple pathology detection is enabled as a set of binary (SVM) classification tasks on the representative histograms  

A sequential series of 114 PA supine chest films obtained at emergency room were analyzed for the presence of any pathology and separately for the presence of specific pathologies such as pleural effusion, cardiac/mediastinal enlargement, and lung opacities. Radiological report obtained by a consensus of two radiologists served as gold standard. Combination of computerized analysis for all pathologies resulted in detection of 38 out of 46 abnormal chest films (sensitivity =83%), specificity 91%, negative/positive predictive values 89/86% respectively. Sensitivity for detection of each specific pathology was low (≤59%) with high specificity 96-100%.  

We showed that the use of the BoW model for automatic detection of abnormal chest film is feasible. Our preliminary experience in a clinical setup is promising with reasonable sensitivity and excellent specificity and negative predictive value. Such software might be used in the future for prioritization of urgent cases in overcharged hospitals, thus improving medical treatment.

Sharon, M, Avni, U, Goldberger, J, Konen, E, Greenspan, H, A Novel Automated Software for Identification of Abnormal Chest Films Using the ”Bag of Words” Model.  Radiological Society of North America 2010 Scientific Assembly and Annual Meeting, November 28 - December 3, 2010 ,Chicago IL. http://archive.rsna.org/2010/9008635.html