An Artificial Intelligence Framework for Analysis of Cognitive Impairment without Prior Knowledge
- Nov. 2020
- by Hyunwoo Oh et. al.
We propose a method that can analyze cognitive ability using 2D slices of brain MRI selected to be highly related to cognitive impairment by machine-learning algorithm.
METHOD AND MATERIALS
We obtained data from ADNI1 and ADNI2. (Alzheimer's disease [AD] = 143, cognitive normal [CN] = 218, stable mild cognitive impairment [sMCI] = 113, progressive MCI [pMCI] = 41) We used the Inception-V4 as the models' backbone, known to achieve a high performance among various 2D CNNs. The classifier consists of a fully-connected layer whose input is a concatenated vector by backbone's output 1536-dimensional vector, age, sex and input image's slice-number. Our proposed method consists of three phase. First, 2D CNN is trained using 2D slices of brain MRI with all range, named entire-brain model. Second, XGB is trained using probabilities of 2D slices of brain MRI with all range calculated by entire-brain model. We can extract a feature importance of XGB and it means a degree of relationship between a 2D slice of brain MRI and cognitive impairment. Finally, a new 2D CNN is trained using two 2D slices of brain MRI extracted by feature importance of XGB. We named this 2D CNN as important-slice model. We compared performances of important-slice model and entire-brain model in classifying AD, CN, pMCI, and sMCI.
Using XGB, we selected top two coronal slices with highest feature importance. The slices included medial temporal lobe which consists of the hippocampus, entorhinal, and parahippocampal cortex. In entire-brain model, accuracies and AUC were 82.4% and 0.877 for AD versus CN, and 72.6% and 0.707 for sMCI versus pMCI, respectively. In important-slice model, accuracies and AUC were 87.5% and 0.938 for AD versus CN and 80.4% and 0.791 for sMCI versus pMCI, respectively. Visualization of salient region by grad-cam of AD subject were more intense than that of CN subject.
The important-slice model using deep learning and gradient boosting algorithm that we proposed achieved higher performance in classifying AD spectrum disorders, particularly distinguishing pMCI and sMCI, compared to the entire-brain model. The findings of feature importance and Grad-CAM provided localization of pathologic change.
The important-slice model provides high performance in classifying pMCI and sMCI with better understanding of pathophysiology of AD spectrum disorders with findings of feature importance and Grad-CAM.