Feature enhanced speckle reduction in ultrasound images: algorithms for scan modelling, speckle filtering, texture analysis and feature improvement
Thesis DisciplineComputer Science
Degree GrantorUniversity of Canterbury
Degree NameDoctor of Philosophy
This thesis aims to contribute to the advancement of research in the field of ultrasound image analysis by developing several novel algorithms in three key areas: (i) modelling, analysis and validation of synthetic ultrasound images, (ii) characterization and reduction of speckle artifacts, and (iii) enhancement of image features. These three areas have been identified based on existing gaps in research and their importance in advanced ultrasound image analysis frameworks for segmentation and classification.
Synthetic models of ultrasound image formation that can generate noise-free ground truth data with intensity and texture characteristics of real ultrasound images, are valuable for machine learning applications and performance evaluation of speckle reduction techniques. This thesis develops a complete framework for synthetic ultrasound image generation incorporating algorithms for image acquisition, sampling and speckle simulation. The framework allows us to simulate image acquisition in both sector and linear scans with varying axial and lateral resolutions.
Speckle artifacts appear in the form of granular noise in ultrasound images, degrading their diagnostic quality. This thesis presents novel algorithms for speckle reduction while preserving edges, fine details, and contrast of the image. The first despeckling framework presented in the thesis uses a novel application of clustering algorithms based on a transformation to wavelet sub-bands, and is inspired by the success of such methods for synthetic aperture radar imagery. The second despeckling framework uses a modified adaptive Wiener filter along with the Canny edge detection and an enhanced steerable pyramid transformation algorithm. Additionally, a coherence component extraction method is used to enhance the overall texture and edge features even in the darker portions of the image.
The filtering operations used in a majority of speckle reduction methods induce blurring that affects edges and other fine structures necessary for accurate clinical interpretation. Ultrasound images require contrast enhancement as they contain low intensity regions of very low contrast and resolvable details. This thesis proposes a modified version of the contrast limited adaptive histogram equalization specifically designed for ultrasound images, where the optimal clip-limit used by the algorithm is automatically determined based on the quality metrics evaluated from output frames, and the size of the image tiles (contextual regions) is also determined automatically using a global entropy function. The thesis also gives an in-depth analysis of the proposed system using three different target distribution functions, and three interpolation techniques.
An integral part of the experimental analysis phase of every algorithm presented in this thesis is an extensive performance evaluation of the developed framework using both quantitative and subjective assessments of image quality. The synthetic images generated by our algorithm are analysed using second order statistical features, and texture features to compare their image quality against real ultrasound images, and to obtain the optimal sampling resolution and interpolation scheme. Similarly, the outputs of the despeckling frameworks are also analysed using several quantitative metrics of image quality such as structural similarity index and edge preservation index. Since image quality does not often directly correspond to usefulness, subjective evaluation by clinical experts is also given equal importance as quantitative evaluation.
The thesis demonstrates that the proposed framework for synthetic ultrasound image modelling can generate images with texture features closely matching real ultrasound images. The thesis proves the effectiveness of the novel clustering algorithm in wavelet transformed sub-bands for edge preserving speckle suppression. The thesis also establishes the capability of the proposed despeckling algorithm based on steerable pyramid transformation in achieving high speckle reduction levels while enhancing texture and image features. An extensive quantitative analysis of the modified contrast limited adaptive histogram equalization method shows that the proposed algorithm could generate outputs with improved visual quality, contrast, and structural content preservation in ultrasound images, compared to several well-known methods for feature enhancement. Finally, the thesis proposes an integrated framework based on the above methods for feature enhanced speckle reduction in medical ultrasound images.
The thesis also outlines future research directions based on the work presented in the thesis and on-going research work on the development of convolutional neural networks, using ultrasound video frames as input.
Index Terms: Medical ultrasound scans, Synthetic ultrasound image modelling, Speckle simulation, Interpolation, Feature extraction and selection, Gray level co-occurrence matrix, Local binary patterns, Image quality measures, Speckle filtering, Clustering algorithm, Steerable pyramid transform, Contrast limited adaptive histogram equalization.