| Περίληψη: | The liver is an important organ of the human body which performs basic functions related to digestion, metabolism, immunity and nutrients storage in the body. Liver disease can lead to liver failure which, if untreated, leads to death. Therefore, accurate diagnosis for targeted therapy of liver diseases is of great importance. Generally there are two types of liver disease, focal and diffuse. Focal liver disease (or Focal Liver Lesions - FLLs) are considered abnormal liver tissue that is concentrated in a small area, such as a tumor or cyst. They are further classified into benign and malignant. The latter usually result in death and therefore the precise characterization of an FLL as malignant is of great importance. The Diffuse Liver Disease (or Chronic Liver Disease - CLD) is distributed almost uniformly throughout the liver tissue. It causes constant and recurring inflammation of the liver tissue that is replaced by connective tissue (fibrosis), leading progressively to cirrhosis. Cirrhosis is the final stage of the disease, leading to liver failure or hepatocellular carcinoma (HCC), a malignant FLL, and eventually death.
The consequences of the development of liver disease make early and accurate diagnosis very important in order to confront them. Liver biopsy is considered ‘Gold Standard’ to date for accurate diagnosis of liver diseases. Liver Biopsy though, is invasive and thus has potential for complications for the patient and increased cost. Therefore, non-invasive approaches, mainly in the medical imaging field, have been developed and evolved the last few years in order to provide a valid alternative to liver biopsy.
The aim of this thesis was to study and develop novel image processing and image analysis algorithms for Ultrasound and Magnetic Resonance Imaging (MRI) for the differential diagnosis of liver disease so that they can be useful additions to the software of the respective imaging medical equipment. Through these algorithms, the diagnostic performance of Radiologists in hepatic diseases can be significantly improved.
During this thesis, four novel medical image processing and analysis algorithms whose performance surpassed those of the respective approaches of the literature were developed.
In detail, the 1st algorithm evaluates FLLs by using image processing and analysis methods in contrast enhanced ultrasound videos. As a first step, a rough estimation of FLL’s position and borders are initially set by calculating Continuous Wavelet Transform (CWT) coefficients (Maxima and Zero Crossings) on each frame of the CEUS video. Then this border estimation is set as initialization step to a Markov Random Field segmentation algorithm which gives the final accurate FLL border delineation. Through this process, the FLL is tracked by the algorithm through all video frames and its borders are extracted accurately independently of FLL’s position, shape or inner intensity change. For each frame of the video that FLL’s borders are extracted, mean intensity values in the FLL and a small Region of Interest (ROI) outside and near the FLL are calculated and a Time-Intensity Curve (TIC) is extracted. Then TIC features are extracted and are fed to a Support Vector Machine (SVM) classifier for final FLL benign to malignant differentiation
The 2nd and 3rd algorithms evaluate CLD through processing and analysis of Ultrasound Shear Wave Elastography (SWE) images. These algorithms extract the colored region with stiffness values from the B-Mode Grayscale values of the SWE image automatically and proceed to an inverse color (RGB) to stiffness conversion, converting the color map to stiffness map. Then, they analyze all the region of stiffness values (2nd algorithm) and subregions of it (3rd algorithm) and extract textural features that a subset of significant ones, is fed to an SVM classifier differentiate between healthy subjects and patients with CLD.
Finally, the 4th algorithm performs differential diagnosis on FLLs through processing and analysis of Magnetic Resonance Images (MRI). Through CWT Multiscale Analysis and FCM segmentation algorithm, the FLL’s borders are extracted accurately. Then textural and morphological features are extracted, a subset of which, the most significant, are fed to a Probabilistic Neural Network (PNN) classifier for FLL differentiation to Benign, HCC and Metastasis.
As a conclusion, the algorithms that were developed during this thesis have confronted successfully clinical issues that are related to liver disease diagnosis through imaging and when equipped, they can improve the diagnostic accuracy of a medical imaging machine.
|
|---|
