透過您的圖書館登入
IP:44.200.77.59
  • 學位論文

基於機器學習與深度學習之 FDG-PET 腦造影分類模型基於重度憂鬱症之診斷

Using FDG-PET Brain Signals for Diagnosing Major Depressive Disorder Based on Machine Learning and Deep Learning

指導教授 : 陳中平
共同指導教授 : 李正達(Cheng-Ta Li)

摘要


重度憂鬱症是一種容易轉變為慢性、病情易惡化且容易衍發出頑固性的心理疾病,近年的研究也顯示憂鬱症更傾向為腦功能異常的神經疾病。約三分之一的重鬱症患者無法對藥物有良好的反應,這類患者被稱為頑固型憂鬱症,若能在治療前得知就診者是否為憂鬱症以及時否為頑固型憂鬱症便能提早制定更優的治療計畫,節省患者時間及金錢的成本,也能及早使用藥物以外的療法,例如:電刺激療法,提高治癒機會。本研究使用107位健康案例、59位非頑固憂鬱症以及59位頑固憂鬱案例的腦部正子造影(PET)搭配機器學習和深度學習演算法,觀察分析機器學習得出之結果和先人研究之結果異同,並藉由機器學習分類器輔助診斷。 我們將腦造影圖預處理後,使用並比較三種特徵選擇方法決定的輸入模型的訓練特徵,並使用集成學習(ensemble)將六個模型中表現最佳的五個之分類結果投票,最終選出一個答案。除此之外,我們也使用深度捲積神經網路 (CNN),並和傳統機器學習的表現比較以驗證電腦視覺類的深度學習模型在正子造影的表現。本研究中設計了三種環境,分別是 : (1) 分類健康案例與憂鬱案例、(2) 分類非頑固型憂鬱與頑固型憂鬱以及 (3) 分類健康、非頑固以及頑固型案例。最終在獨立的測試集上我們在機器學習得到最佳分類準確度為 : 83%,88%,78%;以及在深度學習模型上 : 79%,88%,74%。

並列摘要


Major depressive disorder (MDD) is a mental disorder which is tend to become chronic, easy to deteriorate and to develop intractable. Recent studies pointed MDD is a neurological disease with abnormal brain function. Approximately one-third of MDD patients didn’t respond to the treatments, and such patients is called treatment-resistant depression (TRD). Patients with TRD could receive more appropriate treatment if we could discriminate their group before treatment, and these will save resource for the patients and increase the probability to remission. In this study, we used PET image from 107 healthy controls (HC), 59 non-TRD (NTR), and 59 TRD patients along with machine learning (ML) and deep leaning (DL) for classifying the subject into 3 groups. After the preprocessing of images, we use three feature selection strategies to decide the input features, and we applied ensemble learning to combine the best 5 results from 6 ML models for the final answer. Besides, we also constructed a convolutional neural network (CNN) for the comparison with conventional ML algorithm and verify the performance of PET image on classifying depressive subjects. We designed 3 classification tasks, which is (1) HC vs MDD, (2) NTR vs TRD, and (3) HC vs NTR vs TRD. Finally, we test the model on an independent testing set and yielded accuracies of 83%, 88%, and 78% for the machine learning models, and accuracies of 79%, 88%, and 74% for the CNN models.

參考文獻


1. Lépine, J.-P. and M. Briley, The increasing burden of depression. Neuropsychiatric disease and treatment, 2011. 7(Suppl 1): p. 3.
2. Murray, C.J., et al., Disability-adjusted life years (DALYs) for 291 diseases and injuries in 21 regions, 1990–2010: a systematic analysis for the Global Burden of Disease Study 2010. The lancet, 2012. 380(9859): p. 2197-2223.
3. Li CT, L.M., Evidence-Based Somatic Treatments and Brain Abnormalities for Suicidality in Patient with Depression. Journal of Suiciology, 2022.
4. Warden, D., et al., The STAR* D Project results: a comprehensive review of findings. Current psychiatry reports, 2007. 9(6): p. 449-459.
5. Papakostas, G.I., Managing partial response or nonresponse: switching, augmentation, and combination strategies for major depressive disorder. The Journal of clinical psychiatry, 2009. 70(suppl 6): p. 11183.

延伸閱讀