中文摘要

在眾多類型的數位醫學影像檢驗中，超音波檢驗是一種最為人接受且不會對人體產生副作用的科技，但是由於超音波醫學影像含有大量的斑點、雜訊及組織紋理，傳統的影像切割技術常常無法得到良好的效果。本計畫結合紋路分析技術及類神經網路模式在超音波影像上對乳房腫瘤的輪廓做切割。在本計畫取得腫瘤紋路特徵方面，我們針對超音波影像多樣化的的特性，找到合適的紋路特徵後，我們將訓練類神經網路並運用分水嶺轉換技術來逼近真正的腫瘤輪廓，傳統的影像切割大都是利用影像的梯度數值，但是在充滿斑點、 雜訊雜訊的超音波影像中，梯度數值常導致錯誤的結果，因此本計畫研究方向是採利用紋路特性來切割腫瘤超音波影像，我們將會充分利用紋路特性於腫瘤切割，以提高腫瘤切割的精確性，使得找到的輪廓可以與專業醫生所認定的輪廓近似。

關鍵詞：超音波影像、紋路、乳房腫瘤切割、分水嶺轉換、類神經網路

Abstract

Due to the ultrasonic examination would not cause any side effect upon human’s body. The ultrasound became the most acceptable procedure for patients in the different types of digital medical image. The ultrasonic image is also an efficient instrument of the clinical physicians to diagnose the nidus at an earlier stage. Unfortunately, the digital ultrasonic image always comprises speckles, noise, and tissue-related textures, most traditional segmentation techniques for ultrasonic images do not perform well. In this project, we will combine texture analysis techniques and the neural network model to segment the breast tumors in the ultrasonic images. There are two main study steps, the texture features finding and the neural network segmentation.
After finding the appropriate texture features, the neural network model will be trained and then used the watershed transform to find the contour of tumor. Traditional segmentation methods often use the gradient in the image, but it does not perform well for ultrasonic image. Thus, we use the textual information to segment tumor in this project. Because of the ultrasonic image full of variety, we use the texture features to find the most similar contour with the physician for the breast ultrasonic image.

Keywords: Ultrasonic Image, Texture, Breast Tumor Segmentation, Watershed Transform, Neural Network