影像細胞單元元素競爭：從超音波影像選取多個目標的新方法

Abstract

本篇論文是利用分水嶺轉換法和區域競爭法而導引出影像細胞單元元素競爭的新方法，目的是想偵測在超音波影像中的腫瘤邊界。使用多重解析（multi-scale）的高斯平滑化來對影像邊界保留平滑。Sobel濾波來形成梯度映射（gradient map）。最後，在梯度映射上用分水嶺轉換法來生成初始的影像細胞單元元素。並依據概似檢定法來對影像細胞單元元素進行合併和分離。整個過程是逐步的合併和分離一個影像細胞單元元素和另一個影像細胞元元素或區域。最後，我們將會以電腦模擬的影像去驗證我們的新方法，並實際應用在台灣大學醫院的臨床超音波影像分割上。兩著可顯示影像細胞單元元素競爭法在超音波影像上的可行性。

A new approach, cell competition, based on watershed transform and region competition is proposed to detect the locations of tumors in ultrasound images. Multi-scale Gaussian filters are used to smooth the ultrasound image and preserve the edge information. Sobel filter are then applied to generate a gradient map. Immersion simulation in watershed transform is used to generate initial cells on the gradient map. Cell-based region merging and splitting are perform on the original ultrasound images based on the likelihood ratio tests of speckle noises. Stepwise merging and splitting of one cell and the other cell or region are proceeded. Thus,region information in cells is incorporated in cell competition. Simulation studies confirm the performance of this new approach. Clinical studies are also performed on the clinical ultrasound images collected at National Taiwan University Hospital. Both studies demonstrate the practical feasibility of cell competition for ultrasound images.