運用深度學習之三維卷積模型與視覺化預測阿茲海默症

Predicting Alzheimer's Disease Using Deep Learning Three-Dimensional Convolutional Models and Visualization

宋庭溱(Ting-Chen Sung) ； 許敦韋(Tun-Wei Hsu)

《中華放射線技術學雜誌》 46卷2期 (2022/06) Pp. 114-121

摘要

阿茲海默症（Alzheimer's Disease, AD）是一種不可逆的退化性神經退行性疾病。臨床上對於阿茲海默症的診斷，主要是以較為主觀的量表評估，尚未有一個確切的量化標準，且在疾病初期尚無法準確診斷。隨著電腦科技的進步以及深度學習的發展，卷積神經網絡（Convolutional Neural Network, CNN）在磁振造影（Magnetic Resonance Imaging, MRI）影像以及其他醫學影像進行AD診斷的圖像識別方面取得了很大進展。然而，由於這些成像數據的可用性有限，有效地將CNN用於AD診斷仍然是一個挑戰。透過建立在3D卷積自動分類器的基礎上，並經過預訓練，可以捕獲結構性MRI掃描中的解剖形狀變化並運用3D CNN來預測AD的MRI結構變化，提供客觀觀點作為醫師及早診斷的參考。我們也採用CAM（Class Activation Map）的可視化來解釋3D CNN模型在識別磁振造影T1結構像中的深層特徵。我們可以清楚的看到3D CNN模型關注影像的那些結構，而得到的這個分類結果。最後在Open Access Series of Imaging Studies（OASIS）提供的資料庫數據上，大於60歲以上的共有235人，其中100人為輕度到重度AD，我們保留30%當作測試資料評估此模型效能，經由訓練資料結果所產生模型經由驗證此模型，所得到的平均準確率有74.04%。利用3D CNN來分類診斷，並應用了可視化方法來了解模型的加權的行為，以對大腦MRI結構影像進行分類。它可以擴展到理解CNN的行為，用於其他疾病診斷系統。

關鍵字

深度學習 ； 三維卷積神經網絡 ； 磁振造影 ； 阿茲海默症

並列摘要

Alzheimer's disease (AD) is an irreversible degenerative neurodegenerative disease. The clinical diagnosis of Alzheimer's disease is mainly based on subjective scales, and there is no quantitative standard for the diagnosis. With the advancement of computer technology and the development of deep learning, Convolutional Neural Network (CNN) has made great progress in image recognition for AD diagnosis using Magnetic Resonance Imaging (MRI) and other medical system. However, due to the limited availability of these imaging data, the effective use of CNNs for AD diagnosis remains a challenge. By building on the 3D convolutional auto-classifier and pre-training, it is possible to capture anatomical shape changes in structural MRI scans and apply 3D CNN to predict MRI structural changes in AD, providing objective data to inform physicians for early diagnosis. We also adopt CAM (Class Activation Map) visualization to explain the deep features of 3D CNN model in identifying the structural image of T1 in MRI. We can clearly see that the 3D CNN model focuses on those structures of the image and obtains this classification result. Finally, A total of 235 people over 60 years of age, 100 had mild to severe AD, were retained as 30% of the test data to evaluate the model's efficacy, and the model was validated with an average accuracy rate of 74.04% on the OASIS (Open Access Series of Imaging Studies database). The 3D CNN was used to classify the diagnosis and a visualization approach was applied to understand the behavior of the model weighted to classify the structural brain MRI images. It can be extended to understand the behavior of the CNN for other disease diagnostic systems.

並列關鍵字

Deep learning ； 3D convolutional neural networks ； magnetic resonance imaging ； Alzheimer's disease