全自動乳房超音波之腫瘤偵測

Abstract

乳癌是目前女性最常發生的癌症之一，在全世界乳癌是女性十大致死率排名第一位。藉由檢測及提早治療方式能有效地減少乳癌造成的死亡。在許多乳房檢測儀器中，全乳房超音波是一個相當重要的影像檢查方法，與乳房X光攝影互補，藉由這兩種方式檢查乳癌是否已發生。然而，手動的超音波對於許多醫生來說是很費時費力的。因此近幾年來，自動的全乳房超音波已經有不少論文研究，希望藉由全自動方式來節省醫師的時間。因為一個病例可掃出數百張影像，醫師必須逐張檢查會花許多時間來診斷。因此，此論文提出一個電腦輔助的乳房腫瘤偵測系統來檢測是否存在可疑區域，以節省醫生診斷的時間。而在此系統中，首先必須作影像前處理: 移除左右兩側黑邊及降低解析度可以減少偵測時間; S 型濾波器可以強化影像對比值; Anisotropic可以降低影像雜訊。再來是腫瘤偵測部份，三維模糊C群平均值法會將影像的灰階值來做分群，疑似腫瘤的部份會被歸類為同一群;再來用連通單元標示演算法將每一個腫瘤裡的像素各別地群聚成連通單元。最後是腫瘤評判標準，根據這些連通單元的特性，去計算每個連通單元中的特徵值再依據評判標準決定是否為可疑腫瘤。在這個實驗中，總共有130個病歷將其分為兩群，將55個病歷為一群當做訓練樣本而其餘75個病歷為另一群做為測試樣本，目前腫瘤偵測準確率約有八成四一，希望在未來是個對醫師診斷有所幫助的輔助工具。

Breast cancer has topped the first spot of women’s cancer occurrence in recent years. The early detection and improved treatment are significant to reduce death of breast cancer. Breast ultrasound is a very important complementary imaging modality with mammography to detect breast cancer. However, the physician needs a lot of time to screen a patient by the manual ultrasound. Recently, the automatic whole breast ultrasound system has been developed to reduce the physician’s time for screening the breast. Because a large number of images are obtained for a case, the physician will still take a lot of time to diagnosis. Therefore, a computer-aided tumor detection system is proposed to find suspicious regions of tumors for reducing the diagnosis time. In the tumor detection system, the images are firstly pre-processed by removing black regions and sub-sampling to reduce the detecting time, then the anisotropic diffusion is applied to smooth the images, and the sigmoid filter is applied to enhance the contrast between tumor and normal tissue. In the detection stage, the fuzzy c-means clustering classifies all the pixels into several groups with similar grey intensities. The connected components labeling is used to find all the connected areas in the 3D ABUS image based on the results of the fuzzy c-means clustering. The tumor criteria are proposed to reduce the number of suspicious regions and only the high potential tumor regions could be remained after applying the criteria. In this experiment, there are 130 cases and they were separated into two groups, the 55 cases are used as the training cases and the remaining 75 cases are used as for testing. By experimental results, almost all tumors can be found by this system and the sensitivity is up to 84.1% (79/94) with 2.37 FP rate per case. The proposed tumor detection system is a good tool to help the diagnosis of doctors.