Dorota Duda, Marek Krętowski and Johanne Bézy-Wendling
In this work, a system for the classification of liver dynamic contest- enhanced CT images is presented. The system simultaneously analyzes the images with the same slice location, corresponding to three typical acquisition moments (without contrast, arterial- and portal phase of contrast propagation). At first, the texture features are extracted separately for each acquisition mo- ment. Afterwards, they are united in one “multiphase” vector, characterizing a triplet of textures. The work focuses on finding the most appropriate features that characterize a multi-image texture. At the beginning, the features which are unstable and dependent on ROI size are eliminated. Then, a small subset of remaining features is selected in order to guarantee the best possible classification accuracy. In total, 9 extraction methods were used, and 61 features were calculated for each of three acquisition moments. 1511 texture triplets, corresponding to 4 hepatic tissue classes were recognized (hepatocellular carcinoma, cholangiocarcinoma, cirrhotic, and normal). As a classifier, an adaptive boosting algorithm with a C4.5 tree was used. Experiments show that a small set of 12 features is able to ensure classification accuracy exceeding 90%, while all of the 183 features provide an accuracy rate of 88.94%.
A new approach to the liver segmentation from 3D images is presented and compared to the existing methods in terms of quality and speed of segmentation. The proposed technique is based on 3D deformable model (active surface) combining boundary and region information. The segmentation quality is comparable to the existing methods but the proposed technique is significantly faster. The experimental evaluation was performed on clinical datasets (both MRI and CT), representing typical as well as more challenging to segment liver shapes.