本論文主要針對阿茲海默症患者與健康人之間的臉部不對稱性進行詳細比對與分析。為此，我們透過相機擷取每位受試者1200張影像，照片經過前處理篩選出最正面照片，並由MediaPipe提取468個三維臉部特徵點。為了修正因相機位置及人臉大小所造成的影響，將x、y座標進行正規化處理。同時，為了減少因人臉偏移角度所產生的不對稱性，我們使用普氏分析產生對位誤差，並將這些對位誤差作為後續分析的特徵。研究對象包括137位健康人與137位阿茲海默症患者。
在分析過程中，我們比較了左右半臉各220個特徵點的對位誤差。經T-test檢驗後，我們發現，在220個對位誤差特徵中，患者的數值顯著大於健康人，僅有19個對位誤差特徵的數值小於健康人。這一結果顯然支持了我們關於阿茲海默症患者臉部不對稱性較高的假說。
根據上述計算得出的特徵，我們利用多層感知器（MLP）演算法進行分類，得到了如下分類性能指標：準確率（accuracy）為0.666，敏感度（sensitivity）為0.652，特異性（specificity）為0.684，陽性預測值（positive predictive value）為0.7143，陰性預測值（negative predictive value）為0.619，及馬修斯相關係數（MCC）為0.3349。此外，我們還計劃在未來的研究中，透過生成臉部網格（Face Mesh）和時間序列分類（Time-Series Classification）進一步提升分類的準確性與穩定性。

This paper primarily focuses on comparing and analyzing facial asymmetry between Alzheimer's patients and healthy individuals. To achieve this, we captured 1200 images of each subject using a camera. After preprocessing to select the most frontal photos, we extracted 468 three-dimensional facial landmarks using MediaPipe. To correct for variations caused by camera position and face size, the x and y coordinates were normalized. Additionally, to reduce asymmetry caused by facial tilt, we used Procrustes analysis to generate alignment errors, which were subsequently used as features for further analysis. The study included 137 healthy individuals and 137 Alzheimer's patients.
During the analysis, we compared the alignment errors of 220 facial landmarks on both the left and right halves of the face. After performing a T-test, we found that in 220 alignment error features, the values for patients were significantly higher than those for healthy individuals, with only 19 alignment error features being lower in patients. This result clearly supports our hypothesis that Alzheimer's patients exhibit higher facial asymmetry.


Based on the computed features, we utilized a Multi-Layer Perceptron (MLP) algorithm for classification, achieving the following performance metrics: accuracy of 0.666, sensitivity of 0.652, specificity of 0.684, positive predictive value of 0.7143, negative predictive value of 0.619, and Matthews correlation coefficient (MCC) of 0.3349. Furthermore, we plan to enhance classification accuracy and stability in future research by generating face meshes and using time-series classification.

論文審定書 i
誌謝 ii
摘要 iii
Abstract iv
目錄 vi
圖目錄 ix
表目錄 xi
1 第一章 緒論 1
1.1 前言 1
1.2 研究動機與目的 2
1.3 文獻回顧 3
1.4 論文架構 5
2 第二章 實驗方法 6
2.1 實驗資料來源前處理 6
2.2 設備規格 7
2.3 實驗流程 9
3 第三章 影像前處理、產生臉部特徵點與正規化 12
3.1 影像前處理 12
3.2 產生臉部特徵點 12
3.3 正規化 15
4 第四章 各分析方式篩選受測者資料 16
4.1 分析方法介紹 16
4.1.1 普氏對位誤差特徵 16
4.1.2 距離誤差特徵 17
4.1.3 深度誤差特徵 17
4.2 最佳照片篩選 18
4.2.1 普氏分析對位誤差 18
4.2.2 深度誤差 19
4.2.3 距離中值誤差 20
4.2.4深度誤差篩選最佳照片 20
5 第五章 普氏分析特徵統計學結果 21
5.1 普氏分析特徵統計學比較 21
6 第六章 影像處理及資料擴增 24
6.1 全局調整因子C 24
6.2 照片處理及擴增 24
6.2.1 亮度調整 25
6.2.2 RGB 調整 26
6.2.3 對比度調整 26
7 第七章 機器學習模型及訓練結果 28
7.1輸入特徵 28
7.2機器學習模型 28
7.2.1 模型架構及參數 28
7.2.2 機器學習模型資料分類 29
7.3 運用投票方式增加精度 30
7.3.1 直接投票 30
7.3.2 刪除部分照片後重新投票 31
8 第八章 結論與未來展望 33
8.1 結論 33
8.2 未來展望 33
9 參考文獻 34

[1] Y.-H. Yang, R. Liscic, and J. Dominguez, "Framework of treating Alzheimer’s dementia," Brain Science Advances, vol. 5, no. 2, pp. 82-93, 2019.
[2] " Taiwan Dementia Association." http://www.tada2002.org.tw/About/IsntDementia.
[3] "WHO."
https://www.who.int/news-room/fact-sheets/detail/the-top-10-causes-of- death"
[4] P. Hammond et al., "Face–brain asymmetry in autism spectrum disorders," Molecular psychiatry, vol. 13, no. 6, pp. 614-623, 2008.
[5] S. Balestrini et al., "Increased facial asymmetry in focal epilepsies associated with unilateral lesions," Brain Communications, vol. 3, no. 2, p. fcab068, 2021.
[6] R. S. Marcucio, N. M. Young, D. Hu, and B. Hallgrimsson, "Mechanisms that underlie co‐variation of the brain and face," genesis, vol. 49, no. 4, pp. 177-189, 2011.
[7] O. E. Linden, J. K. He, C. S. Morrison, S. R. Sullivan, and H. O. Taylor, "The relationship between age and facial asymmetry," Plastic and Reconstructive Surgery, vol. 142, no. 5, pp. 1145-1152, 2018.
[8] J. Kim et al., "Numerical Approach to Facial Palsy Using a Novel Registration Method with 3D Facial Landmark," Sensors, vol. 22, no. 17, p. 6636, 2022.
[9] J. Barbosa, W.-K. Seo, and J. Kang, "paraFaceTest: an ensemble of regression tree-based facial features extraction for efficient facial paralysis classification," BMC Medical Imaging, vol. 19, no. 1, pp. 1-14, 2019.
[10] G. S. Parra-Dominguez, R. E. Sanchez-Yanez, and C. H. Garcia-Capulin, "Facial paralysis detection on images using key point analysis," Applied Sciences, vol. 11, no. 5, p. 2435, 2021.
[11] Y. Xia, C. Nduka, R. Y. Kannan, E. Pescarini, J. E. Berner, and H. Yu, "AFLFP: A Database With Annotated Facial Landmarks for Facial Palsy," IEEE Transactions on Computational Social Systems, 2022.
[12] S. E. O’Bryant et al., "Staging dementia using Clinical Dementia Rating Scale Sum of Boxes scores: a Texas Alzheimer's research consortium study," Archives of neurology, vol. 65, no. 8, pp. 1091-1095, 2008.
[13] "Intel Stereo Camera D435i." https://developers.google.com/mediapipe/solutions/vision/face_landmarker .
[14] "MediaPipe Face landmark detection guide." https://developers.google.com/mediapipe/solutions/vision/face_landmarker.

[15] J. C. Gower, "Generalized procrustes analysis," Psychometrika, vol. 40, pp. 33-51, 1975.
[16] Taud, H., & Mas, J. F. (2018). Multilayer perceptron (MLP). Geomatic approaches for modeling land change scenarios, 451-455