電腦輔助腮腺超音波乾燥症檢測系統開發

Computer Aided Diagnosis for Sjögren's Syndrome in Sonographic Parotid Gland Imaging

乾燥症為免疫風濕科常見之慢性、進行性自體免疫疾病，主要侵犯唾液腺及淚腺，而造成病患口乾舌燥。腮腺超音波檢查被發現具有潛力協助診斷乾燥症及追蹤腮腺受侵犯之嚴重度，根據腮腺超音波影像質地不均勻的程度及結構的改變予以判讀。以肉眼來判讀病患之腮腺超音波是否出現乾燥症特徵的方法，存在著不同判讀者可能出現診斷結果不一致的問題。因此如何更客觀正確地檢測出腮腺超音波之乾燥症特徵，並建構出診斷乾燥症之檢測系統，為數位化輔助診斷一重要議題。本論文提出電腦輔助腮腺超音波之乾燥症診斷架構，將腮腺超音波影像經醫師以互動方式圈選關注區域(Region of Interest, ROI)，經紋理分析、空間域與頻率域特徵萃取，並以資料探勘方式篩選出有利於腮腺乾燥症分類之特徵，最後以分類器檢測乾燥症，測試各種特徵組合對乾燥症檢測的績效。特徵萃取在紋理分析之共同發生矩陣中，採用θ=(0°,45°.90°.135°）、d=(l,2,5,7)參數，再結合空間域與頻率域資訊，共計180特徵。績效衡量之指標包括操作特徵(Receiver Operative Characteristic, ROC)曲線之Az值及正確率。以Az值為指標，績效衡量之模型包括全部特徵於支援向量機(Support Vector Machine, SVM)分類器、經循序前向搜尋(Sequential Forward Search, SFS)選擇特徵於SVM分類器、由分類與迴歸樹(Classification and Regression Tree,C&RT)選擇特徵並於SVM分類。實驗結果顯示以SFS選擇64個特徵建構SVM分類器為(Az=0.9131)優於使用全部180個特徵建構SVM分類器(Az=0.8765)及C&RT選擇14個特徵建構SVM分類器(Az=0.8742)。以正確率為指標，績效衡量之模型包括前述三 者及C&RT決策樹。實驗結果顯示以SFS選擇64個特徵建構SVM（正確率90.42％）優於使用全部180個特徵建構SVM分類器（正確率81.67％）、以C&RT選擇14個特徵建構SVM分類器（正確率80.83％）及C&RT決策樹（正確率79.17％）。SFS結合SVM分類器方式執行資料探勘，優於以全部的特徵為輸入向量、2經C&RT選擇14個特徵建構SVM分類器、以及以C&RT決策樹等分類績效。採用SFS資料探勘並結合SVM分類器，減少輸入向量資料量，並同時提昇檢測績效。

In Sjögren's syndrome, the parotid gland is subject to inflammatory disease of autoimmunity. Sonography of the parotid gland is a common tool for diagnosis of Sjögren's syndrome. A diseased parotid gland has a characteristic texture in the sonographic image. The major drawback of manual inspection is inconsistency in the quality of the diagnosis. A computerized diagnostic tool based on objective criteria is needed to improve the inter-observer consistency. In this paper, a computer aided diagnostic system is proposed to detect Sjögren's syndrome. The regions of interest in the sonographic image are interactively selected by users. Features from texture analysis, spatial and spectral domains are then derived. A data mining technique is applied to choose the meaning features. Then, the features are fed into the classifier for Sjögren's syndrome classification. The performance of various feature combinations is compared. 180 features extracted from the co-occurrence matrix under the parameters of θ=(O°,45°, 90°, 135°) and d=(l,3,5,7), the spatial and the frequency domains. The performance indices include the Az value of the ROC curve and the accuracy rate. For Az value, the models included the Support Vector Machine (SVM) with the input 180 features, the features selected by Sequential forward selection (SFS). and the features selected by Classification & Regression Tree (C&RT). In terms of the Az value, SVM input with the 64 features selected by SFS and classified (Az=0.9193) performs best. The second is SVM with the all 180 features(Az=0.8765). The third is the SVM with the 14 featured selected by C&RT (Az=0.8742). In terms of the accuracy rate, SYM with the 64 features selected by SFS (Accuracy Rate =90.42%) performs best. The second is the SVM with the all 180 features (Accuracy rate 81.67%). The third is the SVM with the features selected by C&RT (Accuracy Rate =80.83%). The last is the feature selected and classified by C&RT (Accuracy Rate =79.17%). Experimental results show that in comparison to other combinations, SFS feature selection and SVM classification yields the best results in sonographic pattern classification of Sjögren's syndrome. SFS simultaneously increases performance and reduces system complexity.