| 摘要: | 本論文以類神經網路發展腫瘤診斷之技術，期望能縮短系統檢測腫瘤之時間與判斷是否有腫瘤及其型態(如大小、位置與光學係數等)。類神經網路建立是以有限元素法前向運算，模擬單一近紅外光光源入射乳房，經組織散射與吸收後，於乳房周緣將呈現之光強度分布，令無腫瘤與有腫瘤之周緣光強度曲線相減，進而從光強度差異曲線中選取相關特徵值(AD, AW1, AW2, α)，再經正規化後作為類神經網路輸入神經元的輸入值；而其相對應之腫瘤型態(contrast, radius, distance, radiation angle)，亦經正規化後作為輸出神經元的輸出值，藉此訓練階段建立類神經網路神經元間的權重值。最後，由實驗、模擬及人體量測狀況模擬三種數據，測試類神經網路診斷方法，並顯示以下結果：(1)本方法對於腫瘤大小及方位診斷之準確度高，而對於腫瘤光學係數診斷之準確度，相對於大小及方位較低，但腫瘤大小及方位足以作為腫瘤診斷之基準，故本方法具可靠性，並可應用此估測做為腫瘤斷面成像之初始值；(2)藉此測試亦顯示腫瘤參數contrast、radius對於特徵值AD及AW1具有特性相近效應，使得類神經網路輸入參數及輸出參數之間呈現一對多關係，此關係將作為後續課題之研究；(3)針對人體量測可能遇到之狀況做模擬，顯示儀器光源強度不同無造成診斷誤差，因此可忽略此變因；而背景光學係數改變及加入乳腺因子會造成診斷誤差，故如何降低誤差為後續必須考量之課題。 This thesis presents an artificial neural network (ANN) technique for breast tumor diagnosis expected to shorten measurement time and discriminate tumor categories. The method is design with four inputs (AD, AW1, AW2, α) passing through the artificial neural network (ANN) to obtain the tumor categories (contrast, radius, distance, radiation angle). The inputs represent the characteristics from the difference between the measured intensities of the inhomogeneous and the homogeneous phantom, and then the ANN is training by simulation data, which is simulated by finite element forward method. Finally testing our method by simulation and experiment data, then we have three concluded points: (1) Tumor location and radius can be estimated more precisely than tumor contrast. Though tumor contrast is estimated false, using tumor location and radius to diagnosis breast tumor is enough. Because of above we can promise our ANN diagnosis method, and then the method also can be expected to be the initial guess for the inverse solution in the numerical simulation. (2) Contrast and radius have similar relation for AD and AW1, and the relation is possible to cause cross-talk for contrast and radius. This remains under investigation. (3) Changing source intensity can not cause diagnosis error, but changing optical properties of background and adding mammary gland model can cause error. Therefore, how to reduce diagnosis error for above factors is an important problem. |
