|Ph.D Student||Rotman Shira|
|Subject||On Multi-Component Approaches to Diagnostic Ultrasound|
|Department||Department of Electrical and Computer Engineering||Supervisor||ASSOCIATE PROF. Moshe Porat|
|Full Thesis text|
In this research, a component-based modelling approach is proposed for medical ultrasound images processing. According to this approach, the image data may be de-composed into several components, such as diagnostic Regions-Of-Interest (ROIs), speckle pattern and clutter noise. Each component is distinctly processed in order to improve tasks such as denoising, enhancement, segmentation and compression.
The proposed approach is motivated by the significant challenges posed due to the accumulation of enormous volumes of medical images data in general, and ultrasound images in particular. Our study is focused on medical ultrasound imaging, whose main advantages are its non-hazardous radiation, real-time applicability and low cost.
We first present a method for simultaneous image compression and de-speckling, based on the optimization of a coding scheme in the sense of compression rate and reconstruction distortion. The optimization is applied in the rate-distortion domain, where the distortion of the reconstructed images is evaluated with respect to a-priori known de-speckled images. The distortion measure is chosen as a combination of mathematical metrics that reflect the visual similarity of the images as well as their edge preserve properties. The method's performance is demonstrated on ultrasound breast scans and fetal monitoring, where the resulting images quality is assessed using both mathematical criteria, as well as expert radiologists’ scorings. The experimental results demonstrate the proposed method’s ability to efficiently de-speckle medical ultrasound images, whilst compressing them and preserving their diagnostic data.
Furthermore, we introduce an adaptive RF raw data compression scheme, motivated by the high data rate requirements for real-time medical ultrasound imaging systems on one hand, and by the interfering presence of image clutter noise on the other hand. Due to real-time system constraints regarding the data rate flow from the front-end to the back-end, data compression is essential. In the proposed scheme, the quantization parameters of the coding transform are chosen in a manner that produces optimal fidelity with respect to de-cluttered B-mode clinical and simulated images. The performance of the proposed compression scheme is demonstrated for clinical and software-simulated images. The obtained results show the ability of the proposed scheme to compress the raw data whilst yielding clutter reduced B-mode images.
Moreover, a method for improved B-mode image compression is proposed, where it is assumed that the strong reflectors pixels in the image consist of the main diagnostic information in the image, i.e. comprise the Region-of-Interest (ROI). In the proposed scheme, identification and separation of the ROI pixels is first applied, followed by a Markov model-based inpainting or, alternatively, Shape-Adaptive Transform coding, as well as a lossy compression of the background image. Experimental results of this method for clinical images demonstrate its ability to efficiently compress the images whilst preserving their diagnostic value.
To conclude, the applicability of the proposed approach is assessed on clinical data, taken from fetal monitoring, breast screening and echo-cardiac imaging. The results are analyzed using fidelity criteria and expert radiologists scoring. It is shown that the performance of the proposed methods is comparable to that of state-of-the-art compression methods and de-noising schemes.