- 學位論文
應用對抗生成網路進行資料擴增以提升卷積神經網絡分析正子斷層掃描於輔助阿兹海默症診斷之效果
GAN-based Brain Image Augmentation with 18F-FDG PET for CNN Classification in Alzheimer’s Disease
摘要
在現今社會，已有許多老年人換上阿兹海默症(Alzheimer’s Disease, AD)。阿兹海默症是一種不可逆的神經退化性疾病，會造成老年人喪失記憶、語言障礙、冷漠等問題,因此在早期階段偵測出阿兹海默症是一個重要的議題。正子斷層造影(Positron Emission Tomography, PET)是一種醫學影像技術，搭配氟化去氧葡萄糖試劑(fluorodeoxyglucose, FDG)能得知大腦内的葡萄糖代謝率，以此輔助阿兹海默症的診斷。近期有大量的研究利用深度學習技術分析正子斷層影像，但往往因爲影像資料量不夠大導致深度學習演算法無法良好的發揮，或是造成過擬合的現象。然而在其他醫學領域，已有不少研究利用許多資料擴增的技術合成影像，其中，對抗生成網路(Generative Adversarial Network, GAN)是2014年以來相當流行的深度學習技術，其可以利用模擬影像的統計分佈製造模擬真實資料的合成影像。 在這篇研究中，資料集被按照比例分成訓練資料、測試資料以及一個獨立的資料集，後兩者會被應用於驗證，而訓練資料會被用於卷積神經網絡(Convolution Neural Network, CNN)模型和對抗生成網路的學習。只有用訓練資料作學習的卷積神經網路模型會被當作比較標準，在測試資料集和獨立資料集上分別得到了75.00%和77.50%的準確率。隨後，非條件式的深度捲積生成對抗網路(Deep Convolutional Generative Adversarial Network, DCGAN)和條件式的輔助分類生成對抗網路(Auxiliary Classifier Generative Adversarial Network, ACGAN) 會學習製造合成影像，這兩者生成的合成影像會被分別加入訓練資料，用作進階的卷積神經網路學習，分別為DCGAN-CNN模型和ACGGAN-CNN模型。最終在測試資料集上，兩個進階的模型分別得到了87.50%和88.75%的準確率，但在獨立資料集上，只有ACGAN-CNN模型同在測試資料上取得了大幅度的進步，準確率為92.50%。考量到準確性和穩定性，輔助分類生成對抗網路產生的合成資料更能幫助卷積神經網絡對於阿兹海默症的診斷，因此ACGAN-CNN模型被用作最終的架構。 本篇論文的主要貢獻在於，建立了一個深度學習模型以用於同時診斷阿兹海默症以及輕度認知障礙(Mild Cognitive Impairment, MCI)，不僅在診斷流程的效率上比起過去研究更高，也在各個指標上都取得了更好的表現。除此之外，本篇研究是第一篇應用對抗生成網路模型來解決FDG-PET影像資料不足問題的研究，該模型能達成資料擴增，在未來的醫學領域之深度學習模型訓練上有所幫助。
並列摘要
In this era, a lot of old people are suffering from Alzheimer’s Disease (AD), which is an irreversible neurodegenerative disease and may cause decline in memory. Therefore, AD diagnosis at early stage is vital for preventing deterioration. FDG-PET is a functional neuroimaging that can detect AD by the information of glucose metabolism in brain. Massive Deep Learning framework have aided AD diagnosis recently by analyzing FDG-PET images. However, the powerful ability of Deep Learning framework is based on great amount of data, which is the drawback of existing studies. To solve the shortage of medical image, some researches implemented Generative Adversarial Networks (GAN) on generating synthetic images for data augmentation. In this thesis, CNN models for classify “NC”, “MCI” and “AD” subjects at the same time are training to aid AD diagnosis. Deep Convolutional Generative Network (DCGAN) and Auxiliary Classifier Generative Adversarial Network (ACGAN) are implemented to synthesize FDG-PET images. The synthetic images generated by 2 GAN models are separately added into training set for advanced training as DCGAN-CNN and ACGAN-CNN. The accuracy of baseline CNN model with only real data is 75.00% whereas DCGAN-CNN and ACGAN-CNN model are 87.5% and 88.75% on test set. But ACGAN-CNN keep better improvement on independent set. Considering accuracy and stability, ACGAN and CNN model are chosen as the proposed framework. Our main contribution is that, our CNN model can diagnose MCI and AD in a more efficient way and have reached better performance and stability than previous researches. Besides, it is the first research implementing GAN on FDG-PET image augmentation, which can help stabilize and enhance Deep Learning model in the future.
參考文獻
延伸閱讀
- Karasawa, H. (2018). 以深層3D捲積神經網路診斷阿茲海默症之研究 [master's thesis, National Chiao Tung University]. Airiti Library. https://www.airitilibrary.com/Article/Detail?DocID=U0030-0205201911023185
- 王賀（2020）。整合基因型資訊與生物路徑和基因表現資料提升阿茲海默症預測準確度〔碩士論文，國立臺灣大學〕。華藝線上圖書館。https://doi.org/10.6342/NTU202000208
- Kao, K. C. (2020). 卷積神經網路於正子斷層掃描影像重建之應用 [master's thesis, National Taiwan University]. Airiti Library. https://doi.org/10.6342/NTU202003975
- 康淵、王俊傑、張永鵬、丁鏞、朱明輝（2005）。Integrated Evaluation in Diagnosis for Multiple Faults by Using Discrete Fuzzy Neural Networks。中原學報，33(3)，403-413。https://doi.org/10.6358/JCYU.200509.0403
- Lin, W. Y. (2019). FDG-PET study of brain changes in neuropathic and cancer pain [doctoral dissertation, National Taiwan University]. Airiti Library. https://doi.org/10.6342/NTU201901698