在臨床診斷中，醫師對於超音波影像中的可疑物件利用定義於ACR BI-RADS 的描述項目描述特徵，並藉此估計惡意的可能性，在本研究計畫中，將針對描述特徵的所有特徵，包含形狀、方向、邊緣、邊界、回聲特徵以及後方回聲特徵提出數學化的模式量化為電腦化的BI-RADS 特徵，並期望電腦化的特徵具有與醫師的特徵描述高度的相關性。
腫瘤的形狀與方向特徵，將以最適橢圓作為比較基準來估計腫瘤形狀的不規則程度以及方向，利用地形圖則可以捕捉腫瘤邊緣的針狀與角狀特徵，並且沿著腫瘤的輪廓的兩側分別切割出腫瘤的外緣以及周圍組織，透過比較這兩個區域的回聲特徵可以估計腫瘤邊界的清晰程度，腫瘤的回聲特徵則是以比較腫瘤內回聲特徵的對比描述，最後，腫瘤後方區域被定義並與腫瘤進行比較來觀察聲波通過腫瘤時的衰減程度。
當BI-RADS 特徵被有效的量化，藉由電腦化BI-RADS 特徵的提示可以幫助醫生進行特徵描述工作，電腦化BI-RADS 特徵也可以用來比較可疑物件在不同影像切割方法中的分割效果與醫師感知的相關性，藉此評估影像分割方法的效果，對於使用電腦化BI-RADS 的電腦輔助診斷系統，則可分別提示醫師個別與所有特徵惡意的可能性，協助醫師的評估工作。
In the clinical diagnosis, the sonographic characteristics of a suspicious object within the ultrasound image are characterized by the physicans using the standarized descriptive items defined in the mass category of ACR BI-RADS and then the likelihood of malignancy is estimated according to the descriptions of characteristics. The standarized descriptive items including shape, orientation, margion, lesion boundary, echo pattern and posterior acorstic characteristic classes are mathematically quantified as coputerized BI-RADS features in this work. The interobserver reliability between the descriptions of physicans and the computerized BI-RADS features is majorly concerned.
For the mass shape and orientation characteristics, the best-fit ellipse is used as base line to roughly estimated the degrees of irregular and parallel. Using the distance map, the angular and spiculated characteristics on mass margin are counted. Furthermore, the contour lines on the distance map can be used to partition two ring areas, surrounding tissue and fringe of mass, along the bothside of mass boundary. The difference between these two areas is then utilized to estimate the degree of abrupt interface on lesion boundary. The echo pattern characteristic of a mass is evaluated by the contrast of gray intensities within the mass. An area behind the mass is defined to evaluate the decline degree when the sound passing through the mass as the posterior acoustic characteristic.
When the descriptive items of BI-RADS are effectively quantized as computerized features, the consistencies between different physicans for characterizing a suspicious object in the clinical diagnosis could be improved by the suggestion of computerized BI-RADS features. The reliabilities between the desired objects of physicans and the partition results of different image segmentation methods are also evaluated by the computerized BI-RADS features. By the evaluations, the performance of different image segmentation methods could be decided. Besides, the likelihoods of malignancy evaluated by individual feature or by all features are estimated when the computerized BI-RADS features is utilized to construct the CAD system.
