查詢結果分析
來源資料
頁籤選單縮合
題名 | 主成份分析及逐步選擇法在乳房腫瘤偵測上的應用=Mass Detection in Mammography Using Principle Component Analysis and Stepwise Selection |
---|---|
作者 | 廖炳松; 郭淑美; 于南書; 陳正義; 李三剛; 張建禕; Liao, Ping-Sung; Guo, Shu-Mei; Yu, Nan-Sue; Chen, Cheng-Yi; Lee, San-Kan; Chang, Chein-I; |
期刊 | 中華放射線醫學雜誌 |
出版日期 | 20061200 |
卷期 | 31:6 民95.12 |
頁次 | 頁275-287 |
分類號 | 416.226 |
語文 | chi |
關鍵詞 | 乳房X光攝影; 腫瘤偵測; 主成份分析; 逐步分析法; 主成份特徵轉換; 類神經網路; Mammography; Mass detection; Principle component analysis; PCA; Stepwise selection; SS; PCA transformation; Neural network; |
中文摘要 | 本論文提出一個結合主成份分析及逐步選擇法的乳房腫瘤偵測系統。其相關的乳房組織之特徵空間,是由碎形維度、緊密度、灰階值統計圖的動量、空間灰階值相關統計值、紋路頻譜及紋路特徵編碼等六大類型所組成。 本論文從主成份分析的能量觀點,探索原有特徵空間中212個特徵變數之間的可能線性組合,或是利用主成份分析挖掘少數幾個原始特徵變數來代替全部特徵變數。實驗結果證實主成份分析可以合理地決定獨立特徵變數的有效維度;而且主成份分析加上逐步選擇法可以粹取出有效特徵變數。 在效能因子表現上,針對上述有效特徵變數進行腫瘤偵測,實驗結果驗證機率類神經網路表現相當突出,對於脂肪類、脂肪腺體類及緻密腺體類乳房組織的正確分類率分別為98.4%、89.6%,及87.8%。 |
英文摘要 | This paper aims at constructing a feasible computer-aided mass detection system in mammography. To effectively distinguish mass tissues from normal tissues, the tissue characteristics are represented by a set of features, including fractal dimension, compactness, gray level histogram moments, statistics of spatial gray level dependence, texture spectrum and texture feature coding method, respectively. With reference to the energy viewpoint of principle component analysis (PCA) , we explore possible linear combinations of the raw 212 feature variables, or from the original feature space we will extract several feature variables whose data distribution is complete enough to model that of the original feature space. Experimental results show that PCA is a useful tool to restrict the size of the independent feature variables, and stepwise selection approach in conjunction with PCA can sieve effective feature variables out from the raw feature space. In particular, on applying probabilistic neural network on the filtered feature vectors it has achieved reasonably good performance. The correct classification rates of fatty tissue, fatty-glandular and dense-glandular are 98.4%, 89.6% and 87.8%, respectively. |
本系統之摘要資訊系依該期刊論文摘要之資訊為主。