Development of pattern recognition algorithms for chronic liver disease staging using shear wave elastographic images

• Main Author: Καυκαλέτος, Αθανάσιος
• Other Authors: Καγκάδης, Γεώργιος
• Format: Thesis
• Language: English
• Published: 2020
• Subjects: Pattern recognition; Chronic liver disease; Shear wave elastography; Αναγνώριση προτύπων; Χρόνια ηπατική νόσος; Ελαστογραφία

Chronic Liver Disease is considered a very common cause of death and a significant worldwide burden. Distinguishing between healthy and diseased individuals is crucial for patient management and various diagnosis techniques have been developed to achieve this goal. Shear Wave Elastography is a non-invasive, comfortable and safely repeatable such technique, that displays the stiffness characteristics of the liver parenchyma. Based on the Shear Wave Elastographic images, the staging of Chronic Liver Disease is possible. In this thesis, the process of staging is automated with the use of computer algorithms. Firstly, the textural features of the images were extracted and arranged into arrays called feature vectors. Since the images were taken from subjects of known condition, the feature vector extracted from every image had to be labelled and associated to each category that the image belongs to. The discriminative power of these textural features was evaluated through the process of statistical analysis, more specifically, stepwise regression and only the ones fulfilling the set criteria were kept and used as input for the next part, that of classification. The classification is handled by pattern recognition algorithms that were developed and trained based on the labelled feature vectors. The algorithms that were developed are • Minimum Distance • Least Square Minimum Distance • k Nearest Neighbor • Probabilistic Neural Network • Support Vector Machine These algorithms were validated using a technique called Leave One Out Cross-Validation, a k-Fold method with k equal to the number of the observations, and with the Hold Out Validation method, with a random 70% of the observations used as train-data and a 30% used as test-data. Finally, the performance of these algorithms was evaluated and compared with each other based on the metrics accuracy, Area Under the Curve from ROC analysis and confusion matrices. The highest performance was achieved from the SVM with an accuracy of 84.2% which when compared to relating, recently published work is ordinary and expected.