| Popis: |
Automatic segmentation of liver is a difficult task and moreover, segmenting tumor from liver adds extra dimensionality of difficulty. It is very unfavorable to segment the liver and tumor from abdominal Computed Tomography (CT) images solely based on gray levels or shape alone because of the overlap in intensity and variability in position and shape of soft tissues. To deal with these issues, this thesis proposes a more efficient method of liver and tumor segmentation from CT images using Gabor Features (GF) and three different machine learning algorithms: Random forest (RF), support vector machine (SVM), and Deep Neural Network (DNN). The texture information provided by GF is expected to be uniform and consistent across multiple slices of the same organ. In this thesis, first, an array of Gabor filters is used to extract pixel level features. Secondly, liver segmentation is performed to extract liver from abdominal CT image using three different classifiers: RF, SVM, and DNN trained on GF. Finally, tumor segmentation is done on the segmented liver image using GF and same set of classifiers. The Gabor filter is similar to perception in the human visual system (HVS), and all the mentioned algorithm for classification are robust and accurate ML techniques that have been applied for pixel-wise segmentation.Thirty-one CT image slices were used to validate our proposed method. The 3D-IRCADb (3D Image Reconstruction for Comparison of Algorithm Database) was the source for CT image slices. The classification accuracy for liver segmentation was 99.55%, 97.88%, and 98.13% for RF, SVM, and DNN respectively, while the classifier accuracy for tumor segmentation on the extracted liver segment was 99.52%, 98.07% and 98.45% for RF, SVM, and DNN, respectively. The Dice Similarity Coefficient (DSC) for liver segmentation was 99.03%, 96.79%, and 97.11% for RF, SVM, and DNN, respectively, while the DSC for tumor segmentation on the extracted liver segment was 99.43%, 96.18%, and 95.65% for RF, SVM, and DNN, respectively. |
