Alexandra Nikita MD, Presenter: Nothing to Disclose

Stavroula Mougiakakou, Abstract Co-Author: Nothing to Disclose

Miltiadis Gletsos, Abstract Co-Author: Nothing to Disclose

Georgios Matsopoulos PhD, Abstract Co-Author: Nothing to Disclose

Konstantina Nikita PhD, Abstract Co-Author: Nothing to Disclose

Dimitrios A Kelekis MD, Abstract Co-Author: Nothing to Disclose

To develop and evaluate a Computer-Aided Diagnostic (CAD) system for the classification of hepatic tissue from non-enhanced Computed Tomography (CT) images, into healthy (C1), cyst (C2), hemangeoma (C3), and hepatocellular carcinoma (C4)

CT images with a resolution of 512x512 pixels from both patients and healthy controls have provided input to the system. All hepatic lesions have been validated by needle biopsies, density measurements and the typical pattern of enhancement after the intravenous injection of iodine contrast. The position, size, and extend of the lesions have been defined by an experienced radiologist. A total of 147 free-hand Regions of Interest (ROIs) have been identified 76 of which correspond to C1, 19 to C2, 28 to C3, and the remaining 24 to C4. The above ROI’s are used as input to the CAD system, which consists of two modules: the feature extraction and the classifier modules. The feature extraction module calculates the average grey level, and 48 texture characteristics obtained from the co-occurrence matrices of the ROI’s. The classifier module consists of a feed-forward Neural Network (NN) trained by a novel hybrid method, which uses genetic algorithms (GA’s) to locate a starting point close to the optimal solution, and then the back-propagation algorithm with adaptive learning rate and momentum to refine the NN configuration with local search. This novel hybrid training method is used for the automatic selection of the most robust features, and the automatic adjustment of the NN architecture leading to optimised classification performance of the CAD system

The CAD system could detect the 97.5% of the lesions, while the sensitivity and specificity were 100% and 100% for both C1 and C2, 96.4% and 99.2% for C3, and 91.7% and 99.2% for C4, using only 18 texture features. The execution time required for the classification of one ROI was of the order of a few seconds

The use of texture features, combined with NN techniques can reveal hidden image information. Additionally, the use of a hybrid training method can result in reduction of the CAD complexity, and enhancement of its performance

Nikita, A, Mougiakakou, S, Gletsos, M, Matsopoulos, G, Nikita, K, Kelekis, D, Computer-aided Diagnosis System for the Characterization of Focal Lesions in Hepatic CT Images.  Radiological Society of North America 2004 Scientific Assembly and Annual Meeting, November 28 - December 3, 2004 ,Chicago IL. http://archive.rsna.org/2004/4406290.html