- 學位論文
全域數位乳房攝影之微鈣化群自動偵測系統
Automated Computer-Aided Detection for Clusters of Microcalcifications on Full-Field Digital Mammography
摘要
在國內乳癌是婦女常見的癌症，而乳房微鈣化組織為乳癌發生初期的一個重要特徵。若能及早偵測出乳房微鈣化組織，並分析這些微鈣化實為有效防治乳癌的方法，其中乳房攝影術為主要用來篩檢乳房微鈣化的方法，目前全域數位式乳房攝影術逐漸取代傳統軟片式乳房攝影術，亦有利於影像處理。因此，本研究目的為透過醫學影像處理之技術，於全域數位式乳房攝影影像上偵測出微鈣化組織及微鈣化群，並發展一套自動化的乳房微鈣化群電腦輔助偵測系統。 本研究偵測微鈣化群的步驟主要為影像前級處理、影像增強、微鈣化組織分割、微鈣化分群、以及兩次的降低誤判類神經網路。在影像前級處理使用Log反轉換來對原始影像進行轉換。在影像增強部分是利用小波轉換的技術來突顯出影像上微鈣化的區域，接著選取保留最亮區域的像素來分割疑似的微鈣化，並擷取其形態及紋理特徵作為第一次類神經網路辨識的輸入來減少微鈣化的誤判。最後則將這些疑似的微鈣化分群，同時擷取每群的形態及紋理特徵，作為第二次類神經網路辨識的輸入以減少微鈣化群的誤判，進而找出微鈣化群的位置。而本研究所使用之影像有16張，其中6張為不含任何微鈣化之正常影像，另外10張則為含有微鈣化及微鈣化群之影像，並透過專業醫師圈選出共有20個微鈣化群，且由574個微鈣化組織所構成。 在偵測微鈣化群效能評估的部分，本研究建立之系統可達到靈敏度為100%時，每張影像只發生微鈣化群誤判的個數為1.7個。另外，比較只使用降低微鈣化誤判類神經網路與只使用降低微鈣化群誤判類神經網路的系統偵測效能，其偵測效能皆能達靈敏度為100%，而每張影像發生微鈣化群誤判的個數分別為5個與5.1個。因此，本系統同時使用兩次的降低誤判類神經網路來識別微鈣化及微鈣化群，確實能夠降低誤判的個數，以提升系統的偵測效能。 本研究已建立一套全域數位乳房攝影之微鈣化群自動偵測系統，能夠準確地偵測出微鈣化群的位置，並且有效地降低誤判的個數，因此不但提供給醫師判讀時的參考依據，並能提升判讀時的效率，進而增進醫療品質。
並列摘要
Breast cancer has becoming one of the leading cancers to women in Taiwan. Micro-calcification is one of the early sign for breast cancer. The screen of micro-calcifications mainly depends on mammography. Currently, full-field digital mammography (FFDM) has taking the place of screen-film mammography (SFM) gradually and it is advantageous to image processing. The purpose of this study is to develop a computer-aided detection (CAD) system to identify micro-calcification clusters automatically on FFDM. The CAD system includes six stages which included: preprocessing; image enhancement; segmentation of suspicious micro-calcifications; false-positive (FP) reduction of micro-calcifications; clustering of micro-calcifications; and FP reduction of micro-calcification clusters. At preprocessing stage, the inverted log transform function was used to transform the raw FFDM. Then, this study enhanced the image contrast by using wavelet transform and preserved the brightest areas of the enhanced image as the locations of suspicious micro-calcifications. At the stage of FP reduction of suspicious micro-calcifications, the first artificial neural network (ANN) was used to identify micro-calcifications with the morphology and textural features from each suspicious micro-calcification. After micro-calcifications clustering, the second ANN identification with the features from each cluster was used to reduce FP of micro-calcification clusters. A data set of 16 images was collected, of which 6 images were normal and other 10 images contained 20 clusters with 490 individual micro-calcifications. All micro-calcifications and clusters were marked by experienced radiologist. For performance evaluation, the purpose CAD system can achieve a sensitivity of 100% with 1.7 FP clusters/image. The performance of the CAD system using the first ANN only and using second ANN only also compared, it was found that the CAD system can achieve a sensitivity of 100% with 5 and 5.1 FP clusters/image, respectively. Consequently, using the ANN twice to reduce FP of micro-calcifications and clusters was able to improve effectively the performance of this system. An automated CAD system for clusters of micro-calcifications which uses the raw FFDM as input was developed. This system has the potential to detect micro-calcification clusters with an acceptable sensitivity and low false positives. The use of the CAD system as “second reader” is considered to be one of the promising approaches that may help radiologists improve the diagnostic efficiency.