透過您的圖書館登入
IP:18.225.35.81
  • 學位論文

由腦磁振影像輔助診斷輕度認知障礙與阿茲海默症及定位其顯著影響區域的基於注意力機制之深度多示例學習系統

Attention-Based Deep Multiple Instance Learning for Diagnosis Aid and Discriminative Patch Localization of Mild Cognitive Impairment and Alzheimer’s Disease from Brain MRIs

指導教授 : 鄭士康
共同指導教授 : 張玉玲

摘要


阿茲海默症是一個無法治癒的神經退化性疾病。,在阿茲海默症的患者在逐漸老化的過程中,會出現短期記憶障礙、語言障礙、情緒不穩等等的認知行為問題。,通常阿茲海默症的原因不明,初期的症狀與正常老化相似,因此必須依靠認知測驗、腦部影像以及血液採檢等多項檢驗判斷。,早期的發現以及治療,能延續病患的健康狀態。,本論文提出一個利用腦部MRI影像MRI來輔助阿茲海默症以及輕度認知障礙的診斷的系統:,希望透過影像的自動辨識,廣泛的應用於臨床診療上;,可以初步篩選有潛在問題的病患,再透過專家進一步鑑定,來達成早期偵測及治療的目的。近年來深度學習已經廣泛地被使用在MRI分類,以及影像內容分割問題中。,傳統的機器學習以及手刻特徵,也慢慢被深度學習所取代。,本論文提出一基於注意力機制的多示例深度學習系統,設計一個end-to-end的模型,進行自動化的影像測試辨識。,利用注意力機制的多示例學習,找出每一個區域對疾病的敏感程度,最後生成這一位病患最終的診斷結果。,本篇論文使用一公開資料集ADNI以及一臺灣本土的記憶性認知障礙資料集,分別訓練認知功能正常與阿茲海默症,以及認知功能正常與輕度認知障礙,兩種工作的診斷模型;,並且透過注意力機制,標示出的重要區域探討疾病及各區域的關聯性。,實驗結果顯示本論文在ADNI-1的認知功能正常與阿茲海默症分類工作,上可以達到與目前state-of-the-art相近的結果(Average accuracy=0.85)。,值得一提的是,目前的其他研究,都先抓取出在解剖學上連續的方塊,或者預先定義與疾病高度相關的關注區域(ROI),作為模型的輸入。,我們本研究是目前學界中,首位次不使用任何先前知識來切割3D方塊體積單元(voxel)的論文探討。,如此我們能利用數據驅動的方式,同時訓練特徵以及其重要性分數。,另外在輕度認知功能障礙及認知功能正常的診斷(相較於認知功能正常的病人)上,本論文亦可達到可與state-of-the-art相近的結果(Average accuracy=0.70)。,由於輕度認知功能障礙相對來說是更具挑戰性的問題,我們所提出的模型,未來需要更多訓練資料作學習,使其能夠達成實用的目的。

並列摘要


Magnetic Resonance Imaging (MRI) has been used widely for computer-aided diagnosis in medical image analysis. Recently many works apply deep learning in anatomical or tumor segmentation and disease classification, because deep learning has gradually outperformed many conventional machine learning methods using pre-defined features. The state-of-the-art methods use deep learning methods to automatically extract anatomical landmarks from MRI instead of hand-craft features, and design an end-to-end model for classification and locating the discriminative patch. In this paper, we proposed an attention-based deep multiple instance learning model, which simplifies the anatomical landmark discovery and leverage the attention mechanism to locate discriminative patch and make the model interpretable. Moreover, we cooperate with Psychologists to validate our performance in locating discriminative patch. We evaluate our proposed model in public datasets ADNI-1 and the early amnestic MCI dataset provided by Prof. Y.L. Chang, demonstrating comparable performance with state-of-the-art and reliability of discriminative path on early MCI.

參考文獻


[1] Folstein, M.F., S.E. Folstein, and P.R.J.J.o.p.r. McHugh, “Mini-mental state”: a practical method for grading the cognitive state of patients for the clinician. 1975. 12(3): p. 189-198.
[2] Morris, J.C.J.N., The Clinical Dementia Rating (CDR): current version and scoring rules. 1993.
[3] Nasreddine, Z.S., et al., The Montreal Cognitive Assessment, MoCA: a brief screening tool for mild cognitive impairment. 2005. 53(4): p. 695-699.
[4] National, G.A.U., Dementia: Assessment, management and support for people living with dementia and their carers. 2018.
[5] Chang, Y.L., et al., Regional cingulum disruption, not gray matter atrophy, detects cognitive changes in amnestic mild cognitive impairment subtypes. J Alzheimers Dis, 2015. 44(1): p. 125-38.

延伸閱讀