4.4. 支援向量機實驗
4.4.4. 比較放射狀基礎函數網路與支援向量機實驗結果
整理上面的表格,將放射狀基礎函數網路與支援向量機之實驗的正確 率結果,綜合整理如表 4-13及圖 4-10:
特徵數
分類器 3 4 5 6 7 34(all)
MI 67.50% 70.00% 72.50% 72.50% 71.25%
RBF
F-score 67.50% 68.75% 67.50% 70.00% 68.75% 72.50%
MI 77.50% 72.50% 68.75% 68.75% 68.75%
SVM
F-score 77.50% 75.00% 68.75% 72.50% 71.25% 66.25%
表 4-13、放射狀基礎函數網路與支援向量機分類之結果比較
60.00%
62.00%
64.00%
66.00%
68.00%
70.00%
72.00%
74.00%
76.00%
78.00%
80.00%
3 4 5 6 7 ALL
特徵數
正確率
RBF+MI RBF+F-score SVM+MI SVM+F-score
圖 4-10、放射狀基礎函數網路與支援向量機分類之結果比較
使用放射狀基礎函數網路,使用之特徵選取法不管是最大交互訊息法
還是F-score,分類的正確率結果較平均,並且使用最大交互訊息法的最
高正確率與未進行特徵擷取之正確率相同,而使用支援向量機,最大交 互訊息法與F-score的結果除了前3個分類之正確率最高,以及前 5個分 類正確率相同之外,其他F-score都比最大交互訊息法來的高。而使用全 部的特徵值以支援向量機分類,分類正確率會下降,顯示放射狀基礎網 路在不同特徵數的表現上較為平均,而較多的特徵數則會些許降低支援 向量機的分辨率。
將 放 射 狀 基 礎 函 數 網 路 與 支 援 向 量 機 的 各 類 別 分 辨 率 整 理 在 表
4-14。可以看出支援向量機除了在第 1 類發炎的分辨率 85.71%,與第 3 類鈣化的 68%,低於最大交互訊息法與放射狀基礎函數網路的組合中的 88.57%與 76.00%。但第 4類正常樣本的正確率,60%遠大於使用放射狀 基礎函數網路與最大交互訊息法的 0%,以及 F-score 法的 30%。因此支 援向量機除了整體分辨率較放射狀基礎函數網路高之外,還更具有一般 性。
類別
分類器 1 2 3 4
MI 88.57% 80.00% 76.00% 0.00%
RBF
F-score 82.86% 80.00% 64.00% 30.00%
MI 85.71% 90.00% 68.00% 60.00%
SVM
F-score 85.71% 90.00% 68.00% 60.00%
表 4-14、放射狀基礎函數網路與支援向量機的各類別分辨率比較
5. 結論
本論文之主要目的是探討能夠使用於偵測肩關節旋轉肌肌群病變之 輔助疾病診斷方法。希望可以增進超音波肌腱影像之診斷效能,能夠造 福肩部疼痛的病患並可節省醫院診查之成本。
本研究分別使用了紋路特徵編碼法,灰階明亮度相互關係矩陣,紋路 頻譜,統計特徵矩陣,以及碎型維度等方法,擷取肩關節影像特徵,並 分別利用交互訊息法與F-score法選取有效特徵,再以放射狀基礎函數網 路分類器進行分類,並比較支援向量機實驗結果。
實驗結果發現最大交互訊息法在大部分情形下的確能使分辨率提 高,使用放射狀基礎函數網路可以與全部特徵值分類之正確率相同,而 使用支援向量機則可比全部特徵值分類結果較佳。不過由於本實驗使用
之fselect程式結合了F-score法與支援向量機,其所取的特徵值,會經過
支援向量機與網格搜尋演算法驗證。所以最大交互訊息法所選出的特徵 值雖然前3個為最高值與F-score最高值,以及前5個特徵分類結果相同,
而其他則特徵數在支援向量機的表現較F-score差。
最大交互訊息法選取出來的7種特徵值中,其所使用方法,包括灰階 明亮度相互關係矩陣之中的:(6)總和平均(Sum Average)和(7)總和變異
(Sum Variance); 以 及 紋 路 特 徵 編 碼 法 中 的 :(3)平 均 數 值(Mean
Convergence);灰階明亮度相互關係矩陣中的:(2)對比度、(10)差變異
(Difference Variance),(11)差熵值(Difference Entropy),以及(8)和熵(Sum Entropy)。而使用 F-score 法選取前 7 個特徵,選出的分別是灰階明亮度 相互關係矩陣之中的:(7)總和變異(Sum Variance)和(6)總和平均(Sum Average);紋路特徵編碼法中的:(3)平均數值(Mean Convergence)和(4) 碼 變 異 數(Code Variance); 灰 階 明 亮 度 相 互 關 係 矩 陣 中 的(2)對 比 度
(Contrast)、(4)平方和(Sum of Squares)、和(10)差變異(Difference
Variance)。可以看出灰階明亮度關係矩陣及紋路特徵編碼法在分辨灰階
超音波影像皆有不錯的辨別度。
實驗結果發現支援向量機最高分類正確率為77.5%,較放射狀基礎函
數網路72.5%為高,兩者分類結果之最高的正確率相差5%。而且支援向
量機在放射狀基礎網路分類不佳的正常類別,也具有 60%的正確率。不 過支援向量機需要反覆調整參數才能有效提高分類正確率,特徵數的數 目也會影響參數調整及分類結果,而參數的選擇方式卻還沒有公認的有 效方式。放射狀基礎函數網路之參數選擇雖然同樣有許多方式,不過本 實驗使用的方式,僅需決定隱藏層中心數,並從分類結果從中挑選最好 的即可,執行速度比支援向量機快。
使用放射狀基礎函數網路最高72.5%,而支援向量機最高77.5%,使 用放射狀基礎函數網路與支援向量機的實驗結果分辨率均偏低,有可能
是訓練樣本不足造成,同時假如特徵擷取區域之選擇如果不適當的話,
也會對結果有所影響,這都是未來需要改進的地方。
參考文獻
[1] 蘇木春、張孝德,「機器學習:類神經網路模糊系統以及基因演算法 則第二版」,全華科技圖書,第 1-15-20,3-29-31頁,2004。 [2] 莊家和,「紋路特徵編碼法於紋路分析之研究」,碩士論文,南華大學
資訊管理學所,2002。
[3] 洪明輝,「肩關節旋轉肌肌群之超音波影像電腦輔助疾病偵測與量化 分析」,國科會,計畫編號: NSC94-2213-E-251-001,2005
[4] Battiti, R., “Using mutual information for selecting features in supervised neural net training,” IEEE Transactions on Neural Networks, Vol. 5, pp.
537-550, 1994.
[5] Chen, C., Daponte, J., and Fox, M., "Fractal feature analysis and
classification in medical imaging," IEEE Trans. Med. Imag., Vol. 6, pp.
133-142, 1989.
[6] Chang, C.C. and Lin, C.J, “LIBSVM: a Library for Support Vector Machine”, 2001. software available at:
http://www.csie.ntu.edu.tw/~cjlin/libsvm/index.html
[7] Chen, D.R., Chang, R.F., Chen, C.J., Ho, M.F., Kuo, S.J., Hung, S.J., Moon, W.K., "Classification of Breast Ultrasound Images Using Fractal Feature", Clinical Imaging, Vol. 29, No. 4, pp. 235-245, 2005.
[8] Chen, Y.-W. and Lin C.-J., “Combining SVMs with various feature selection strategies.”, “Feature extraction, foundations and applications",
2005. software available at:
http://www.csie.ntu.edu.tw/~cjlin/libsvmtools/fselect/fselect.py
[9] Haralick, R.M., “Statistical and structural approaches to texture,” Proc.
IEEE, Vol.67, pp.784-804, 1979.
[10]Haralick, R.M., Shanmugam, K., and Dinstein, I., “Textural features for image Classification”, IEEE Transaction on Systems, Man, and
Cybernetics, Vol. 3, No.6, pp.610-621, 1973.
[11]Haykin, Simon, Neural Networks:A Comprehensive Foundation, 2nd edition, Prentice Hall, pp.256-308, 1998.
[12]He, D.C., and Wang, L., “Texture features based on texture spectrum”, Pattern Recognition, Vol. 24, pp. 391-399, 1991.
[13]Horng, M. H., Sun, Y. N., and Lin, X. Z., ”Texture feature coding method for classification of liver sonography”, Computerized Medical Imaging and Graphics, 26, pp.33-42, 2002.
[14]Hsu, C.-W., Chang, C.-C., and Lin, C.-J., “A Practical Guide to Support Vector Classification”, Department of Computer Science and Information Engineering National Taiwan University”, 2003.
[15]Kwak, N. and Choi, C.H.,”Input Feature Selection by Mutual Information”, IEEE Transaction on Pattern Analysis and Machine Intelligence, 24(12), pp.1667-1671, 2002.
[16]Luigi Portinale and Lorenza Saitta. “Feature Selection”. No. D14.1, IST Project MiningMart, IST-11993, 2002.
[17]Ou, Yu-Yen, “QuickRBF: An Efficient RBFN Package”. software available at:
http://csie.org/~yien/quickrbf/quickstart.php
[18]Whitney, A., "A direct method of nonparametric measurement selection", IEEE Trans. Computing., vol. 20, pp.1100-1103, 1971.
[19]Wang, Li, HE, D.C., "Texture classification using texture spectrum", Pattern Recognition, Vol. 23, pp.905-910, 1990.
[20]Wu, C. M., and Chen, Y. C., “Statistical Feature Matrix for Texture
Analysis”, CVGIP:Graphical Models and Image Processing, Vol. 54, No.
5, September, pp.407-419, 1992.
附錄
實驗之超音波影像診斷資料分類說明:1:發炎,2:鈣化,3:斷裂,4:正常
病歷號 超音波下診斷 分類
00166922 no evidence of of RCT 1
00275681 complete tear of right rotator cuff 3
00321474 Tendinosis of right supraspinatus tendon 1 00549520 Partial tendon tear of supraspinatus tendon 2
00568114 No evidence of RCT 1
00595277 Tendinosis of supraspinatus tendon, right 1 00628442 Partial thickness tendon tear of right supraspinatus tendon 3 00636724 Partial thickness tear of the right supraspinatus tendon 3 00742874 Tendon tear of the right supraspinatus tendon 3 00917439 Tendinosis of right supraspinatus tendon 1
01191686 No evidence of rotator cuff tear 1
01238678 Tendinosis of bilateral supraspinatus tendons 1
01339440 Right supraspinatus rupture 3
01856971 Right subacromial bursitis 2
01869279 partial tear of right supraspinatus tendon 3 01870940 Tendon partial tear of supraspinatus tendon 3 02251485 modeate tear of right supraspinatus tendon 3 02362331 Partiral tendon tear of supraspinatus tendon 3 02374165 Tendinosis of left supraspinatus tendon 2 02392123 Complete tendon tear of right supraspinatus tendon 3 02601954 moderate tear of right supraspinatus tendon 3 02845114 Partial thickness tear of the left supraspinatus tendon 3 02910002 Tendinosis of the left supraspinatus tendon 1 03152264 Tendinosis of bilateral supraspinatus tendons 2 03658995 Tendon complete tear of the right supraspinatus tendon 3 03789740 Tendinosis of the left supraspinatus tendon 2 03812407 moderate full thickness tear of bilateral supraspinatus tendon 3
04038034 Left supraspinatus complete tear 3
04038445 Tendinosis of supraspinatus tendon 1
04069811 Small full-thickness tendon tear of right supraspinatus tendon 3
05076465 complete tear of left supraspinatus tendon 3
05296423 No evidence ofRCT 1
05876173 full thickness tear of right supraspinatus tendon 3
06175336 no evidence of RCT 1
06177879 Tendinosis of supraspinatus tendon 1
06288275 Tendinosis of right supraspinatus tendon 1 06538802 Tendinosis of right supraspinatus tendon 1 06563322 Tendinosis of left supraspinatus tendon 1
06941632 No evidence of RCT 1
07017131 Tendinosis of bilateralsupraspinatus tendon 1 07048950 Small full-thickness tendon tear in left supraspinatus tendon 3
08073766 Right RCT complete tear 3
08128205 moderate partial thickness tear of left supraspinatus tendon 3 08165711 Tendinosis of left supraspinatus tendon 1 08490326 Tendinosis of right supraspinatus tendon 1 08560764 Negative organic lesion of shoulder in US exam 4 09003750 Tendinosis of bilateral supraspinatus tendon 1 09443491 Tendinosis of supraspinatus tendon, right 1
09492090 No evidence of rotator cuff tear 4
09519028 Suspected muscle atrophy of the left upper arm 1
09623635 No evidence of rotator cuff tear 4
09983016 Tendinosis of supraspinatus tendon 1
10045331 Suspect partial tear of right supraspinatus tendon 2 10126364 Tendinosis of the left supraspinatus tendon 1 10159537 Tendinosis of bilateral supraspinatus tendons 2 10225119 Suggest right supraspinatus small partial tear 3 10423560 Tendinosis of left supraspinatus tendon 1 10587703 Tendon tear(full thickness) of right supraspinatus tendon 3
10604456 Left supraspinatus tendenosis 1
10632187 No evidence of rotator cuff tear 4
10740505 Tendinosis of left supraspinatus tendon 1
10752148 Tendinosis of supraspinatus tendon 1
10784745 suspect partial tear of right supraspinatus tendon 3 10799162 Tendinosis of bilateral supraspinatus tendon 1
10834187 Right subacromial bursitis 3
10888252 Tendinosis of bilateral supraspinatus tendon 1 10907095 Tendinosis of bilateral supraspinatus tendon 1 10917234 Tendinosis of bilateral supraspinatus tendon 1 10929398 Tendinosis of right supraspinatus tendon 1 11125123 Tendinosis of the left supraspinatus tendon 1 11146004 Tendinosis of the right supraspinatus tendon 1
11183099 Tendinosis of supraspinatus tendon 1
11191627 Full thickness tear of right supraspinatus tendon 3