A Multi-scale Feature Extraction FilterFor Breast Lesions

[[abstract]]隨著近幾年來女性乳癌的死亡率快速增加，在乳房損害的偵測過程中，乳房X 光攝影與乳房超音波檢查這兩種影像式的檢測方法，在臨床治療上對於惡性腫瘤 的初期診斷，扮演相當重要的角色。針對電腦輔助診斷系統而言，腫瘤邊緣的幾 何特性通常是可用來辨識良惡性的有意義特徵。為了抽取幾何特徵，傳統的低階 影像處理方式是針對每一種特徵函數，就必須使用一套演算法則來處理判別。這 樣的方式不僅導致時間的耗費，更增加雜訊影響的不可靠度。基於資訊只存在訊 號變動量的觀點，我們使用一種新的處理方式，藉由一微離散週期性小波轉換的 一次處理過程同時取得所有需要的特徵。此外，這種方式可以提供多重解析的特 徵，並在偵測乳房損害過程中達到抗雜訊的效果。論文中所建立的三種基於餘弦 函數所構成的一微離散週期性小波轉換係數，可以表示出在邊緣變化上的一階與 二階導函數特性，並且經由簡單的代數運算可以獲得五種乳癌檢測的辨識特徵。 經過比較發現，Haar 與Daubechies-4 的係數也可以應用在這個的實驗中。在初 步的二十張影像測試中，我們得到具有95%的良惡性辨識率。[[abstract]]Since the female mortality rate of breast caner is fast growing in recent years, breast lesion detection in terms of mammogram and ultrasonic imaging technique has played an important role in clinical medicine for an early diagnosis of the existence of malignant tumors. For a computer aided diagnosis system, the most significant features usually used to discriminate the malignant tumors from benign tumors are the geometric characteristics of tumor’s boundary. To extract the geometric features, the traditional approaches based on low-level image processing technique needs several processes to extract variant features because a feature extraction process generally can be suitable for only one kind of features. This will result in not only time consuming but also unreliability due to the influence of noise. Instead, based on the sense that message may exit just in the variance of a signal, this thesis presents a novel approach to derive all the desired features by taking only one path of the 1-D discrete periodized wavelet transform (DPWT). In addition, the approach can provide multi-scale features which have the property of strong noise resistance for breast lesion detection. Three kinds of cos-based 1-D DPWT filters are designed so that the 1-D DPWT coefficients can denote the first and second derivatives of a boundary shape. Then by associating several simple algebraic computations with the 1-D DPWT coefficients, five kinds of breast lesion detection features can be obtained. For a comparison, the Haar and Daubechies-4 filters are also applied in the experiments. The experimental result of twenty cases shows that the approach can have 95% of success rate of malignant tumor detection