Quantitative Analysis Of Ultrasound Image Features And Differentiation Of Breast Tumors

Ultrasound is routinely used as an adjunct to mammography for detecting and diagnosing breast diseases. With the consistent improvements in ultrasound imaging technology, ultrasound plays important roles in diagnosis of breast lesions because of its characteristics, such as noninvasive, more efficient, relatively inexpensive, and convenient. Especially, ultrasound is useful in detecting breast cancer among women with dense breast tissue which is very common in Chinese women. However, the interpretation of ultrasound images is highly dependent on the experience of the physician who reviews and analyzes the ultrasound images. The lack of uniformity and consensus among observers'use and interpretation of various descriptive terms for malignant and benign lesions often results in inconsistent diagnosis. Computerized schemes can be used to reduce this operator dependency and provide a more objective interpretation of ultrasound findings. The number of unnecessary biopsies of breast lesions may be further reduced with the use of effective computer-aided diagnosis (CAD) system that would improve the specificity of discriminating malignant from benign lesions on breast ultrasound images.The aim of this dissertation is to develop effective measures which can characterize quantitatively the ultrasound image features of breast tumor, to reveal the relationship between the computerized image features and malignancy of breast tumor and rovide the interpretable diagnostic suggestions to physician through the computerized scheme. From the domain knowledge of ultrasound medicine, the features used by physician are summarized and the corresponding computation methods of them are developed. The expected feature measures of breast ultrasound image should be interpretable and in accordance with the perception of humans, and contain the potential diagnostic information. The malignant risk assessment models of breast tumors can be constructed through the statistical learning methods on the collected retrospect samples, which can provide the second opinion to physician.The main contributions of this dissertation include:(1) Segmentation of breast ultrasound image. The completely automatic segmentation of breast tumor on the ultrasound image is rather difficult. Manual segmentation is time consuming and the location of tumor boundary is inaccurate. Especially, the difference of segmentation would influence the reliability of statistics of features computed from the manually outlined tumor contours. Normalized cut (Ncut) and live-wire techniques are applied to automatic and interactive segmentation of breast tumor. Ncut technique readily admits combinations of different clues such as brightness, position and windowed histograms. Incorporating with the proposed priori rules, K-way Ncut can produce the similar result to the manual segmentation. The live-wire method can combine effectively the accurate edge detection and the knowledge of physician. Importantly, the segmentation by live-wire is reproducible that is crucial for the reliable statistics.(2) Breast ultrasound image feature analysis. The image features on shape, margin, boundary, orientation, echo pattern, and posterior acoustic of breast tumor are quantified according to BI-RADS (Breast imaging report and data system) lexicon for breast ultrasound. The shape features have the relatively strong ability to differentiate the malignant and benign tumor from the statistical results. The shape measures of breast tumor are summarized in a synthesis index—shape complexity measure. A rank learning algorithm is proposed to learn the shape complexity measure which is in accordance with the perception of human. From the clinical observation, asymmetry of tumor shape is often a malignant sign which is associated with the tumor growth patterns. The multiscale shape symmetry measures based on local area integral invariant are proposed to quantify the reflective symmetry of breast tumor. The experimental results show that the symmetry between malignant and benign are significantly different and benign breast tumors are more symmetric than malignant ones.(3) Interpretation of breast ultrasound image features. The underlying relationship between computerized ultrasound image features and malignancy of breast tumor might not be linear in nature. Thus, this relationship should be given for the clinical application of the quantitative image features. The decision tree ensemble generated by the cost-sensitive Boosting algorithm is used to approximate the target function of malignancy assessment and reflect qualitatively this relationship. Then, the partial dependence plots are employed to explore and visualize the effect of features on the output of decision tree ensemble. The results provide the visual and qualitative reference to physician, and can be useful to enhance the interpretability of CAD system for breast ultrasound.(4) Malignancy assessment model of breast tumor. The malignant risk assessment models are constructed by statistical learning methods using the quantitative image features as the input variables, which can provide the diagnostic suggestion to physician. The cost-sensitive Boosting algorithm and the maximizing AUC (area under ROC curve) rule ensemble method are proposed to train the models by minimizing the cost-weighted margin loss and maximizing AUC. The sensitivity and specificity of the prediction models can be tuned by these methods. The rule ensemble method can induce the interpretable scoring rules.