以深度學習診斷阿茲海默症之臨床可解釋性研究

Abstract

阿茲海默症 (Alzheimer's disease, AD) 大約佔所有失智症 60% - 70%，它是一種不可逆的神經性退化疾病，目前世界上的治療方法的成效尚不明確，因此患者的病況會隨著時間的進展慢慢的惡化，不僅會影響患者本身，也為家庭帶來沈重的負擔。根據世界衛生組織 (WHO) 的統計，全球失智症人口正逐年攀升，若缺乏一個能快速、準確且大規模應用的輔助診斷工具，現有的臨床醫療體系將面臨巨大挑戰。 本研究旨在發展一個輕量化且具備可解釋性分析的三維卷積神經網路 (3D CNN) 框架，用於分析全腦結構性磁振造影 (structural Magnetic Resonance Imaging, sMRI) 影像，以輔助阿茲海默症的診斷。模型採用輕量化設計，整合了群組卷積 (group convolution)、全域池化 (global pooling) 與高效通道注意力機制 (efficient channel attention)，並以全腦影像作為模型輸入，避免預定義的腦區分割處理，而失去了部分的潛在特徵。在可解釋性分析方面，使用基於遮罩敏感度 (occlusion sensitivity-based) 方法，視覺化模型於分類決策時所關注之關鍵腦區。並透過階層回歸 (hierarchical regression)分析，驗證模型輸出預測值與簡短智能評估量表 (Mini-Mental State Examination, MMSE) 之間的相關性。 結果顯示，本模型僅使用 batch size 為 2 以及模型參數量僅 3.3 百萬 (M)，即於獨立測試集達到 90.6% 的準確率。可解釋性分析結果，模型的分類決策主要依據海馬迴、丘腦以及小腦等與 AD 病理高度相關的腦區。此外，階層式回歸分析顯示，模型輸出預測值之 logit 值與 MMSE 分數間存在顯著負相關性 (β = -0.600, p < 0.001)，即使在控制年齡、性別和教育程度等因素後，仍可有效解釋 MMSE 分數之變異。 本研究所提出之 3D CNN 模型在輕量化的同時，亦具備良好的 AD 分類能力，並透過可解釋性分析，展現其決策過程的透明度與的臨床的應用價值，為臨床提供具有潛力的 AD 輔助診斷工具。

Alzheimer’s disease (AD) is an irreversible neurodegenerative disorder and accounts for approximately 60% to 70% of all dementia cases. The efficacy of current treatments for AD remain unclear. Consequently, the patient's condition progressively deteriorates over time, not only affecting the patient but also imposing a heavy burden on families. According to the World Health Organization (WHO), the global dementia population is increasing annually. Without a rapid, accurate, and scalable auxiliary diagnostic tool, the existing clinical healthcare system will face significant challenges. This study aimed to develop a lightweight and interpretable 3D Convolutional Neural Network (CNN) framework for analyzing whole-brain structural Magnetic Resonance Imaging (sMRI) scans to aid in the diagnosis of AD. The model adopted a lightweight design, and integrated group convolution, global pooling, and efficient channel attention. The model took whole-brain images as input to avoid predefined brain region segmentation processes that might lead to the loss of potential features. For interpretability analysis, an occlusion sensitivity-based method was used to visualize the key brain regions the model focused on during classification decisions. Furthermore, hierarchical regression analysis was employed to examine the correlation between the model's output predictions and Mini-Mental State Examination (MMSE) scores. Results showed that the proposed model achieved an accuracy of 90.6% on an independent test set using a batch size of only 2 and a model parameter count of just 3.3 million. The interpretability analysis revealed that the model's classification decisions primarily relied on brain regions known to associate with AD pathology, such as the hippocampus, thalamus, and cerebellum. Additionally, hierarchical regression analysis indicated a significant negative correlation between the logit values of the model's output predictions and MMSE scores (β = -0.600, p < 0.001). This correlation remained effective in explaining the variance in MMSE scores even after controlling for age, gender, and education level. In conclusion, the 3D CNN model proposed in this study demonstrated excellent AD classification capability while being lightweight and clinically interpretable. The model may serve as a potential auxiliary diagnostic tool for AD.