遷移式學習應用於正顎手術前後對稱性評估

THUIR > College of Engineering > Department of Computer Science and Information > Master's Theses > Item 310901/31837

Please use this identifier to cite or link to this item: http://140.128.103.80:8080/handle/310901/31837

Title:	遷移式學習應用於正顎手術前後對稱性評估
Other Titles:	Using Transfer Learning for Facial Symmetry Assessment Before and After Orthognathic Surgery
Authors:	鄭鈞澤 CHENG,CHUN-TSE
Contributors:	楊朝棟 YANG,CHAO-TUNG 資訊工程學系
Keywords:	遷移式學習;卷積神經網路;深度學習;面部對稱性;資料預處理 Transfer Learning;CNN;Deep Learning;Facial symmetry;Data Pre-process
Date:	2019
Issue Date:	2019-12-16T06:51:33Z (UTC)
Abstract:	遷移式學習是目前影像辨識相關研究中對於小數據集訓練中非常重要的方法，我們將遷移式學習應用於醫療領域，透過卷積神經網路將針對正顎手術前後進行面部對稱性的評分，評分標準的部分我們經由長庚醫院的8位醫師問卷調查，將醫師的評分做為參考標準，但由於面部對稱性評分沒有一定的標準，因此我們將評分做平均並去除1至3個標準差之外的值，讓評分更加一致性，透過處理電腦斷層掃描3D影像將其轉換為等高線圖，此等高線圖保有3D的特徵，由於我們資料量過小所以需要使用資料預處理，我們透過不同的資料預處理方法與資料擴增，將原始資料擴增為100倍，並透過實驗找出最符合我們目標任務的方法，並於本文中使用了四種預訓練模型，本文實驗中比較了四種模型的優劣，而實驗中使用的神經網路為將預訓練模型導入後，增加全連結層以及分類層，並生成隨機變形資料訓練時放入模型，使用此方法以及加入Dropout神經網路層來防止模型過度擬合，由實驗的結果我們最終選擇了Xception模型以及Constant的資料擴增方法達到高達90%的準確率，透過模型預測時產生的信心值，給出模型所預測的評分。本文完成了使用遷移式學習進行訓練，達到針對正顎手術前後的立體顏面影像萃取之等高線圖片進行對稱度的評分，並比較了不同的預訓練模型以及不同預處理方法的優劣，進而找出真實準確率最高的方法。 Assessing facial symmetry is crucial for orthodontists. An accurate appraisement of face symmetry is notable for the development of dentofacial orthopedics diagnosis. To facilitate a successful treatment, an understanding of personal characteristic in the perception of face symmetry become the critical factor. However, there is no standard of facial symmetry score. It depends on the orthodontists expertise in face symmetry judgments. Therefore, it is tough to ensure accuracy. To support the physicians for improving the medical treatments, in this paper, we propose a Convolutional Neural Networks(CNN) with transfer learning method for facial symmetry assessment based on three-dimensional features. In this case, we train the new model through Transfer learning to score facial symmetry. In this work, we convert the computed tomography (CT) scan of a 3D image into a contour map which retains the characteristics of 3D. To find the best result, we use different data pre-processing method and data amplification method. The original data is amplified by 100 times. In our experiment, we compare the quality of the four models, and the neural network architecture used in the experiment is to import the pre-training model. Also, we increase the fully-connected layer and the classification layer. To prevent the model from overfitting, we put the random deformation data during training and Dropout. From the experimental results, we chose the Xception model and the Constant data amplification method to reach up to the accuracy rate of 90%. The score predicted by the model is given by the confidence value predicted by the model.
Appears in Collections:	[Department of Computer Science and Information ] Master's Theses

Files in This Item:

File	Description	Size	Format
107THU00394017-001.pdf		6769Kb	Adobe PDF	174	View/Open

Loading...