CLASSIFICATION MODEL AND INTERPRETABILITY FOR ALZHEIMER'S DISEASE AND BEHAVIORAL VARIANT OF FRONTOTEMPORAL DEMENTIA BASED ON CNN

An improved one-dimensional convolutional neural network (1D-ICNN) model is proposed to diagnose Alzheimer's disease and abnormal frontotemporal dementia. In the model, the output of convolution layer was down sampled and the global average pooling of feature compression was performed. The classification accuracy of this model on brain structure MRI data of 47 patients with Alzheimer's disease and 39 patients with behavioral disorder frontotemporal dementia was 86.63%, which was better than traditional machine learning model and traditional deep learning model. In addition, the SHAP interpretable method was used to interpret the prediction results of the model, and the interpretation results were visualized.