自組性類神經網路應用於乳房X光影像之偵測

Abstract

醫學影像的研究，從早期的X光片、超音波顯像，到現在的核磁共振（MRI）、電腦斷層掃瞄（CT），使得全世界在醫學技術及醫療品質有了大幅的提昇。然而在國內，此等醫療技術的應用仍大量仰賴國外高科技產品的輸入，例如MRI的掃描器，仍是完全倚靠國外的輸入；相關的使用軟體也是購自國外。為了解決此問題，並降低全民醫療成本，實有必要自行發展醫學影像辨認系統。本論文結合圖形識別與類神經網路之演算法則，提出一套改善乳房X光影像的辨認技術，以供乳癌診斷參考。 本論文所提出的概念，主要利用自組性類神經網路（SOM）的演算法則，實施特徵萃取、分類定義及聚類的工作，使乳癌的鈣化組織與腫瘤區塊較準確的顯示出來，並配合影像處理之分析，達到提升辨識速度及診斷準確度的目標。所開發出來的影像處理工具箱，包括影像濾波、SOM特徵擷取與影像邊界描述等。經由本研究的模擬實驗後，在運算複雜度與速度上，均已獲得改善。The progress of medical imaging technologies, from X-ray radiography, ultrasonic sonography to modern age's Magnetic resonance imaging (MRI) and Computed Axial Tomography (CT/CAT) scan has helped the advance of the medical technology as well as the improvement of medical care quality all over the world. However, in Taiwan, the state of art of such technologies are still far behind those of advanced nations. All important medical tools and instruments still rely heavily on the import from other countries, such as Japan and United States. For example, NMR spectroscopy of the MRI machine is still wholly imported from those countries. Even the related soft-wares are also more than 90% purchased from other countries. It is essential to develop our own medical imaging identification systems so as to reduce the future overall medical expense of our country. To contribute to this effort, we propose a new medical imaging technologies, which combines the advanced technologies of pattern recognition and modern numerical methods in neural networking, to improve the power of discretion in analyzing mammography so as to reduce the false rate in the diagnostics of breast cancer. In this study, we applied the methodologies commonly used in computer-aided design systems and self-organizing mapping (SOM) artificial neural networks, such as feature extraction, clustering and filtering, to the ramification of various mammograms. We show that more accurate diagnostics can be achieved with better sensitivity in separating calcified tissues and tumor masses. Higher detection resolution, better recognition efficiency and fast processing speed for mammography are also realized with the aid of new imaging techniques. We also developed an imaging processing toolbox, which contains image filtering, SOM feature extraction and rendering of (blur) image boundary. All newly developed numerical methods and functions (including SOL) can be easily retrieved for image analysis. The result of our computer simulations clearly shows that the complexity of mammography imaging processing algorithm and calculation speed can be significantly improved based on our proposed methods.