Choi & Jin, 2018, Predicting Cognitive Decline with Deep Learning of Brain Metabolism and Amyloid Imaging

# 세줄요약 #

1. 인간에게 치명적인 인지기능 장애를 가져오는 알츠하이머 질병을 진단하기 위하여 CNN 기반의 진단 알고리즘을 개발하였으며, 먼저 AD(Alzheimer Disease)와 NC(Normal Control)을 분류하는 모델을 먼저 학습시킨 후, MCI(Mild Cognitive Inpairment) 환자들이 치매로 전환(Converter or Nonconverter)되는지를 학습시켰다.
2. 학습 데이터로는 ADNI 오픈 데이터셋의 FDG와 AV-45 PET 영상들을 사용하였으며, 이때 딥러닝은 특성 추출을 자동으로 하기에 Spatial Normalization 같이 뇌영상에서 전통적으로 사용되던 영상처리 기법들은 적용하지 않았다.
3. 학습 결과 AD vs NC 환자 분류와 MCI 환자들 중 AD 환자로 전환자/비전환자 분류 모두 새로 사용한 딥러닝 기법이 기존의 특성 추출 & 머신러닝 기법보다 우수한 성능을 나타내었다(정확도 96%, 민감도 93.5%, 특이도 97.8%).

 

 

# 상세리뷰 #

1. Our deep CNN-based approach could accurately predict cognitive decline in MCI patients by combining information of FDG and AV-45 PET images.

• We aimed to develop an automatic image interpretation system based on a deep convolutional neural network (CNN) which can accurately predict future cognitive decline in mild cognitive impairment (MCI) patients using flurodeoxyglucose(FDG) and florbetapir(AV-45) positron emission tomography (PET).

FIGURE 1. Framework for predicting cognitive decline in mild cognitive impairment patients. (A) Deep convolutional neural network architecture is applied to the two PET images, FDG and florbetapir (AV-45). (B) Deep CNN was trained from PET data of AD and NC. 10-fold cross validation was used. After the training, the trained network was directly used for the classification between mild cognitive impairment (MCI) converter and nonconverter.

 

2. For testing whether a MCI subject would convert to AD, baseline PET images without spatial transformation were needed as a feature extraction was automatically performed.

• Accuracy of differentiation between MCI converter and nonconverter (84.2%) outperformed the conventional machine learning approach using SVM with voxelwise feature selection as well as conventional feature-VOI based method.
• ROC comparison results also revealed that the accuracy of prediction was significantly higher than other methods.

FIGURE 2. ROC curves for deep CNN. ROC analyses were performed for the classification of AD (A) and the prediction of MCI conversion (B). ROC curves of feature VOI-based approaches using FDG and AV-45 PET were also drawn

 

Table 2. Sensitivity, specificity and accuracy for the discrimination between AD and normal controls and the prediction for MCI converters

3. Our approach could provide a quantitative variable, ConvScore, to be used as a fusion biomarker for multimodal images.

• ConvScore calculated from baseline PET images of MCI patients was significantly correlated with the longitudinal change of cognitive measurements at 1 year and 3 years.

FIGURE 3. Correlation between output of the network and longitudinal changes of cognitive measurements. The last layer provides an output scores for AD or MCI converter, which are defined as ConvScore. (A-D) ConvScore was significantly correlated with the change of Cognitive Dementia Scaling Sum of Boxes (CDR-SB) (r=0.37, p<0.0001), Alzheimer Disease Assessment Scale-Cognitive Subtest (ADAS-Cog) (r=0.29, p=0.0001), Functional Activities Questionnaire (FAQ) (r=0.40, p<0.0001) and Mini-mental State Examination (MMSE) (r=- 0.30, p<0.0001) from baseline to 1 year follow-up. (E-H) The significant correlation between ConvScore and the change of the measurements from baseline to 3 years was also found (r=0.63, p<0.0001 for CDR-SB; r=0.24, p=0.004 for ADAS-Cog; r=0.67, p<0.0001 for FAQ; r=-0.61, p<0.0001 for MMSE).

 

* Reference: Predicting Cognitive Decline with Deep Learning of Brain Metabolism and Amyloid Imaging.

Hongyoon Choi, Kyong Hwan Jin, Alzheimer’s Disease Neuroimaging Initiative

Behav Brain Res. 2018 Feb 14 Published online 2018 Feb 14. doi: 10.1016/j.bbr.2018.02.017