4D Feature Extraction from Echocardiography Images using Local Phase Information

Abstract

Cardiovascular disease is one of the leading causes of deaths worldwide which has resulted into increased usage of the cardiac imaging. There are a number of cardiac imaging modalities, among which echocardiography is the most widely used because of its fast acquisition, cost effectiveness and harmless nature. However, echocardiography images may also suffer from poor characteristics that include: intensity dropout, low contrast, low signal to noise ratio and speckle appearance. These characteristics make ultrasound images difficult for subsequent processing and analysis and hence an approach independent of contrast, noise and illumination is required. Local phase based approach is suitable for low contrast and noisy images and outperforms the conventional intensity based feature extraction in case of ultrasound images. In this work, we study the extraction of endocardial and epicardial boundaries which have step edge characteristics. For this purpose, we employed FA measure as it tends to extract asymmetric features (i.e. step edges) in an image. The extraction algorithm was developed using monogenic signal, which is a generalization of analytic signal for higher dimensional analysis, to avoid the complexities of oriented filters. Spatial based feature extraction ignores the continuity in the heart cycle and it has been shown that spurious features may be detected. It is believed that spurious features (noise) that are not consistent along the frames can be excluded by considering the time information. We have investigated the spatio-temporal based feature extraction to explore the effect of adding time information in feature extraction process. In this thesis we have studied 4D feature extraction to determine whether addition of time information improves the features extraction and what change it brings by comparing its results with 3D feature extraction. To evaluate the impact of time dimension, an image-driven LV segmentation method was applied to both the 3D and 4D images. The visual and quantitative results showed that 4D FA measure relies on the temporal resolution of images. In fact, 4D does not always produce better results and the results may actually degrade when temporal resolution is poor. However, for frames where the time resolution is good and LV shape does not show fast changes (e.g. start and end of cardiac cycle) with time, 4D FA measure performs better or comparable to 3D FA measure.