Performance evaluation in [18F]Florbetaben brain PET images classification using 3D Convolutional Neural Network.

Abstract

High accuracy has been reported in deep learning classification for amyloid brain scans, an important factor in Alzheimer's disease diagnosis. However, the possibility of overfitting should be considered, as this model is fitted with sample data. Therefore, we created and evaluated an [18F]Florbetaben amyloid brain positron emission tomography (PET) scan classification model with a Dong-A University Hospital (DAUH) dataset based on a convolutional neural network (CNN), and performed external validation with the Alzheimer's Disease Neuroimaging Initiative dataset. Spatial normalization, count normalization, and skull stripping preprocessing were performed on the DAUH and external datasets. However, smoothing was only performed on the external dataset. Three types of models were used, depending on their structure: Inception3D, ResNet3D, and VGG3D. After training with 80% of the DAUH dataset, an appropriate model was selected, and the rest of the DAUH dataset was used for model evaluation. The generalization potential of the selected model was then validated using the external dataset. The accuracy of the model evaluation for Inception3D, ResNet3D, and VGG3D was 95.4%, 92.0%, and 97.7%, and the accuracy of the external validation was 76.7%, 67.1%, and 85.3%, respectively. Inception3D and ResNet3D were retrained with the external dataset; then, the area under the curve was compared to determine the binary classification performance with a significance level of less than 0.05. When external validation was performed again after fine tuning, the performance improved to 15.3%p for Inception3D and 16.9%p for ResNet3D. In [18F]Florbetaben amyloid brain PET scan classification using CNN, the generalization potential can be seen through external validation. When there is a significant difference between the model classification performance and the external validation, changing the model structure or fine tuning the model can help improve the classification performance, and the optimal model can also be found by collaborating through a web-based open platform.