利用超音波技術輔助診斷甲狀腫瘤是目前常見的診斷方法,然而目前對於由超音波取得的影像主要還是由醫療人員以肉眼來進行腫瘤特徵之判斷,此方法對於判斷腫瘤特徵有兩項主要缺點。首先,不同的檢視者對於相同的影像所做出之判斷有不唯一性,其次,由於是以肉眼進行判斷,尚無法以連續化之數值來表示特徵於影像呈現上的程度差異。 文獻上,對於診斷甲狀腺腫瘤之特徵於灰階超音波影像上主要有鈣化、低回音、異質化、邊緣模糊等特徵。而於超音波都卜勒影像上主要有觀察於腫瘤周圍的部分之血液量。針對這些腫瘤特徵,在此研究中,我們嘗試去將每個特徵以數值量化,並比較量化特徵與傳統質化特徵對於甲狀腺腫瘤良惡性之判斷效力。另外,由於將特徵以數值量化,故我們亦將量化特徵以費雪判別分析尋找一最佳線性組合,以期能判別使甲狀腺腫瘤之良惡性之分類最佳化。為驗證本研究所提出之量化特徵之方法,我們利用由台大醫院提供之樣本資料來進行特徵之數值量化,並以台大醫院所提供之病歷資料來做判別準確性之評斷。
The ultrasound technology is widely used in diagnosis of the thyroid nodules. Currently, evaluation of nodule features on the ultrasound images is subjectively made by the medical examiners. There are two major shortcomings of the present approach. First, the result of evaluation varies under different medical examiners. Second, due to the naked-eye evaluation, the evaluation result is usually categorical and is incapable of showing the gravity level of the observed feature. In the literature, the major features for diagnosis of thyroid nodules are microcalcification, hypoechoic lesion, heterogeneous echo texture and margin blur on gray-scale images. In power Doppler images, the major feature is the observation on the volume blood flow of the peripheral area around the nodule. In this research, we attempt to quantify these features on the images. The diagnosis power is compared between the quantitative features and the conventional qualitative features. Furthermore, the Fisher Linear Discrminant is employed on these quantitative features to obtain a linear combination for a better classification power. To validate the proposed approach, we use the data sample provided by National Taiwan University Hospital(NTUH) to retrieve the quantitative features, The clinical data also provided by NTUH, is then used to evaluate the diagnosis accuracy.