在乳房超音波影像之部分區域使用模糊分群法 分析乳房密度

Abstract

根據中華民國行政院國民健康局的統計報告，2008年女性乳癌為女性癌症發生率的首位，死亡率為第四位。且著名期刊“A Cancer Journal for Clinicians”內所載的 “Cancer statistics 2012”指出美國女性乳癌發生率為第一位，死亡率高居第二位。另有許多研究指出乳癌發生率和乳房密度有高度的正相關性，因此如何量化計算分析出乳房的密度資訊成為一門重要的研究方向。 當前乳房密度分析的影像來源常見的有，乳房X光攝影(mammography)、3-D核磁共振掃描(3-D MRI)、以及3-D全乳房自動超音波(3-D automatic whole breast ultrasound, ABUS)等，然而在乳房X光攝影方面，其影像為2-D的資訊，乳房中的腺體脂肪皆被投射在2-D平面上，並不能表示出乳房的立體模型，無法反映出真實的乳房密度，此外有曝露在放射線之下的風險。另外3-D MRI雖可清楚描繪出乳房的完整三維影像，並且反映出真實的乳房密度並且無放射線，但拍攝費用相當昂貴。而超音波影像無侵入性，也無放射線的特質，較被一般民眾接受。因此本篇研究乃使用3-D全乳房自動超音波影像來做乳房密度量化分析。 在3-D超音波影像中將乳房區域與非乳房區域完整分割需要複雜的運算，另外一方面，必須經由多次的3-D全乳房自動超音波影像掃描才足以涵蓋整個乳房的範圍，因此必須將各次掃描的結果組成完整乳房之後才供密度分析使用，但在組成的過程中會因為每次掃描時的擠壓以及掃描所造成的重複區域而使得組合後的影像與真實的乳房有所誤差，因此本篇論文所提出的方法不需要組合多次掃描的影像，而是只取各次掃描的部分區域影像來做密度分析這些區域分別為AP PASS 乳頭的六點鐘及十二點鐘方位的VOI以及LAT PASS乳頭側面方位的VOI，並分別將各區域的結果以及三個區域總和結果與核磁共振掃描影像以及取完整區域的乳房組織的3-D全乳房自動超音波影像所分析出來的密度作比較。 總計有67筆乳房影像分析資料，關於結果的比較我們使用的是相關係數以及線性迴歸分析。而比較的結果在與乳房X光攝影比較方面，三區域總合結果、12點鐘方位VOI、6點鐘方位VOI以及側邊方位的相關係數R2分別為0.630、0.574、0.584、0.611；與完整乳房的3-D超音波影像方面，三區域總合結果、12點鐘方位VOI、6點鐘方位VOI以及側邊方位的相關係數R2分別為0.963、0.925、0.882、0.919；與核磁共振影像方面，三區域總合結果、12點鐘方位VOI、6點鐘方位VOI以及側邊方位的相關係數R2分別為0.832、0.731、0.783、0.819。結果顯示即便在不同的影像來源能有高度相關性，也能有合理的結果作為診斷參考。

According to the Taiwan Cancer Registry (2006-2008) of Bureau of Health Promotion, Department of Health, R.O.C. (Taiwan), the first place of the ten leading cancers for females is breast cancer, and the death rates is at the fourth place. Likewise, in United States breast cancer is the most commonly diagnosed cancer for females and it is in the second position of the top ten deadliest cancers. Therefore, the research for quantization and analysis of breast density is more and more important. In recent years, some studies have suggested that breast density is highly associative with the incidence rates of breast cancer. Mammography, magnetic resonance imaging (MRI), and ultrasound are commonly used for breast cancer examination. However, mammographic images are two-dimension (2-D), the volume is not represented truly and all tissues are overlapped. Consequently, the breast density may be overestimated. The automated breast ultrasound (ABUS) is developed as a screening tool and is used to support the diagnosis of mammography. Therefore, research of breast density analysis with ABUS becomes more and more popular. In ABUS, several passes were needed to scan the whole breast. The multi-pass images are manually merged as a whole breast image by using the nipple as the landmark to align the overlapped regions. However, the shape of the breast might be deformed due to the pressure during the scanning. That is, some regions may be overlapped or lost which cause the error of density estimation in the merged images. Besides, some shadows under the nipple and distortion near the boundary of images which may cause the failure of density estimation for ABUS. In this study, only parts of the automated whole breast image were extracted for density analysis. Three volumes of interest (VOIs) were extracted from the 12, 6 o’clock, and lateral position of nipple. Finally, the fuzzy c-means classifier was applied to differentiate the fibroglandular and fat tissues for density estimation. The results were compared with the results of mammography, whole ABUS, and MRI. The linear regression coefficients between the three positions (12 o’clock VOI, 6 o’clock VOI, and lateral VOI) and mammography are 0.574, 0.584, 0.611 respectively. Averaging the three positions obtained the coefficient of 0.630 for the correlation with mammography. The linear regression coefficients between the three positions (12 o’clock VOI, 6 o’clock VOI, and lateral VOI) and ABUS are 0.925, 0.882, and 0.919, respectively. Averaging the three positions obtained the coefficient of 0.963 for the correlation with ABUS. The linear regression coefficients between the three positions (12 o’clock VOI, 6 o’clock VOI, and lateral VOI) and MRI are 0.731, 0.783, and 0.819, respectively. Averaging the three positions obtained the coefficient of 0.832 for the correlation with MRI