Study On Image Analysis And Processing Methods For Ultrasound Images Of Intracardiac Masses

The intracardiac masses are hazardous to the human health. Generally, they are abnormal structures within or immediately adjacent to the heart, which may cause the obstruction in many organs. The early detection and diagnosis will greatly reduce the risk of cardiac surgery and improve the survival rate. Echocardiography is one of the most widely used medical imaging modalities to diagnose intracardiac masses, for its noninvasive nature, real time, low cost and convenience. It becomes the preferred technique in the cardiac disease diagnosis and cardiac function evaluation clinically.Because of the interferences of ultrasound backscatter signals during the imaging, the ultrasound images suffer from the suboptimal image qualities, including large amount of speckle noise, signal drop-out and missing contours, which leads to the difficulty in the quantitative analysis and the low diagnosis accuracy. Therefore, this dissertation focuses on the hot topics in the ultrasound image analysis to develop an automatic classification system, assisting the radiologists to distinguish between intracardiac tumor and thrombi in the clinical practice.1. For the image denoising, the strong speckle is the main factor affecting the diagnosis accuracy. As the locally-based image despeckling methods fail in retaining edges and structure, a maximum likelihood estimator and non-local means based algorithm is proposed to reduce the speckle while well preserving the tissue details. To decrease the computational complexity, a novel global despeckling approach combining the sparse representation and non local means algorithm is developed. It exploits the repetitive character of structures in the whole image, realizing the truly global despeckling. Considering the temporal correlation of the echocardiogram sequences, a newly method based on the motion estimation is further proposed to extend the despeckling into spatial-temporal aspect. The above three methods are capable of effectively suppressing the speckle noise, laying a solid foundation for the future work.2. In the image segmentation part, due to the complex cardiac structure and tiny masses, the general methods are unsatisfactory in the intracardiac mass segmentation. In order to overcome the difficulties in identifying blurred and missing contours in echocardiogram, a novel supervised learning method based on selective ensemble learning algorithm and active contour model is proposed, which incorporating the shape prior into texture information, achieving a better detection performance. To solve the two key problems in active contour model, including its sensitive to the initial contour and external forces vulnerable to the noise, an automatic segmentation approach based on the sparse representation and the modified active contour model is present. With the initial contour obtained by the sparse representation and the sparse coefficients as new external forces, the segmentation results are similar to the current clinical standard of manual tracing. The experiments demonstrate the high robust and accurate results of our proposed methods for the intended task.3. For each segmented mass, a computerized scheme is designed to extract and quantify nine features from the echocardiogram. Besides the common rules which the radiologists usually refer to, including the mass movement and the base length, more novel texture characteristics are suggested as supplementary. All features are able to distinguish from the intracardiac tumor and thrombi.4. Finally, all nine features are implemented into a sparse representation classifier to assess the overall classification performance. The study compares the affects of different feature subsets and the classification performance with the state-of-the-art classifiers.A complete automatic classification system is established based on all above parts to identify the intracardiac mass in echocardiogram.97clinical echocardiogram videos are collected to assess the efficacy. Our system shows the best performance, achieving an accuracy of96.91%, a sensitivity of100%and a specificity of93.02%, much higher than the result of the manual classification. The results explicates that our system is capable of classifying intracardiac tumor and thrombi in echocardiogram, potentially to assist the radiologists in the clinical practice.