Fully Automatic Segmentation Of Breast Ultra-sound Images

Breast cancer is one of the most often detected kind of cancers which threatens health and life of woman.Early and accurate disease diagnosis plays great role in cancer treatment.Clinical research has shown that if the cancer is detected in early stages it can relatively easily cured without making much harm to a patient.One way to detect breast cancer is to use ultrasound imaging.Medical ultrasound uses sound waves with high frequencies that cannot be heard by humans(>20,000 Hz).The pulses send into tissues and reflected back with different properties are recorded and displayed as an image.It is quite a convenient tool to observe diseases such as tumor and other abnormalities in the human body.In order to reduce mistakes in cancer diagnosis and help doctors,computer itself can find tumor area and surrounding tissue structure which is the main task of this paper.One way to achieve this task is to apply semantic image segmentation.This segmentation method can calculate not only object location but also determine to which class it belongs.The classification model is used to calculate probability distribution for each pixel and then the result can be optimized to make accurate segmentation.This paper first summarizes the research status of image segmentation at home and abroad,and then proposes the idea of the image semantic segmentation method in this paper.The main idea of this research is to apply full convolutional neural network for feature extraction.Initially FCN is trained with training samples and one-hot images and then after feeding another set of images it can create segmentation results.The chosen architecture of the fully convolutional network is U-net and its two other variants.These other two variants called Dual U-net and Tight U-net are based on idea of deep convolutional framelets.This idea has an advantage over the initial U-net architecture because it allows to reduce noise during segmentation process thus giving more accurate result.All three variants are tested during experiments to evaluate which one is better.Though FCN can give quite accurate segmentation result they are still can be refined.This paper proposes two methods to achieve such goals.One of them is calculate wavelet form for every image in the training set – such approach can give neural network more details during training.Another method is to use conditional random field which takes into account the whole context of the image and produces smoothed prediction results removing coarse edges.Experiments shows that proposed methods in this paper can solve segmentation problems even if the size of training set is limited and unclear image features.