以3D卷積神經網路偵測阿玆海默症之全大腦視覺化分析

Abstract

隨著醫療技術發展與衛生條件改善，全球人口逐漸高齡化，罹患失智症的人數不斷攀升。阿茲海默症（Alzheimer's Disease，AD）是目前最常見的失智症類型，尚無有效的治癒療程，僅能仰賴藥物控制。因此，AD的早期診斷及定期追蹤具有臨床重要性，其中又以低侵入性技術最符合風險利益比，例如磁振造影（Magnetic Resonance Imaging，MRI）。此外，多數失智症的影像深度學習研究為了提高分類準確度，經常會加入各式影像前處理、甚至預先切割出海馬迴或其他預設區域，限縮模型學習特定區域的體積或形狀變化等解剖訊息。如此做法雖然能剔除不具訊息的影像，提高模型收斂效率，但似乎違反了深度學習自動尋找特徵的本質。基於上述，本研究提出VGG-based 3D CNN模型：結合三維卷積神經網路（Convolutional Neural Networks，CNN）及VGG架構，改良架構配置，並將全連接層替換為Global Average Pooling。本研究以全大腦MRI影像作為訓練資料集，除了頭皮剝離及影像信號正規化，不做其他影像前處理，藉此驗證深度學習模型從MRI影像學習大腦組織結構特徵之能力，並以Grad-CAM視覺化結果檢視模型判別AD的依據是否符合臨床診斷。實驗結果顯示VGG-based 3D CNN診斷AD的準確度可達87.5%，模型分類主要關注側腦室、腦溝和腦皮質外圍，和臨床的專家經驗與研究相符—證明我們提出的VGG-based 3D CNN模型確實能學習到具有臨床意義的AD大腦影像特徵，具有一定的診斷準確度。值得一提的是，模型在不同資料集個案中的關注結構區域非常相似，這也證明VGG-based 3D CNN模型能驗證不同來源的影像資料集，對於臨床應用更具潛力。

With the rapid aging of the global population, the number of people living with dementia is growing. Alzheimer's disease (AD) is the most common form of dementia. As of now, there is no cure for AD and medication only slows disease progression. Therefore, early diagnosis and regular monitoring of AD are of clinical importance, preferably by using techniques with minimal invasiveness and a reasonable risk-benefit ratio, such as magnetic resonance imaging (MRI). Additionally, most deep-learning-based imaging studies of dementia apply various image preprocessing and some even pre-segment anatomical regions such as hippocampus to improve the accuracy of classification or prediction. Constraining the model to learn structural features such as volume and shape from predetermined regions of interest may improve the converging speed. However, removing regions assumed to be uninformative seems to conflict with the fundamental strength of deep learning. In light of the above, we proposed a VGG-based three-dimensional convolutional neural network (3D CNN) model where the original VGG configuration was modified and the fully-connected layer was replaced with global average pooling. MRI images of the whole cerebrum were used as the training and testing datasets with minimal preprocessing (skull/scalp stripping and signal normalization) to validate the ability of the model to learn the spatial information from MRI images. Then, gradient-weighted class activation mapping (Grad-CAM) was used to visualize whether the model picked up the anatomical features comparable with those previously reported and applied in clinical diagnosis. Our results showed that VGG-based 3D CNN achieved an accuracy of 87.5% for AD diagnosis and mainly focused on the ventricle, sulcus, and surrounding areas of the cortex. It was also noted that our model picked up highly similar anatomical regions in different datasets. In conclusion, the VGG-based 3D CNN model proposed in this study was able to detect AD with a satisfactory accuracy rate based on clinically interpretable anatomical features.