Related Experiment Video
Updated: Jun 20, 2026
09:03
Radiotracer Administration for High Temporal Resolution Positron Emission Tomography of the Human Brain: Application to FDG-fPET
Published on: October 22, 2019
10.3K
Performance evaluation in [18F]Florbetaben brain PET images classification using 3D Convolutional Neural Network
Seung-Yeon Lee1,2, Hyeon Kang2, Jong-Hun Jeong3
1Department of Translational Biomedical Sciences, Dong-A University, Busan, Korea.
Plos One
|October 20, 2021
Summary
Convolutional neural networks (CNNs) show promise for classifying amyloid brain PET scans in Alzheimer's disease diagnosis. External validation revealed overfitting issues, but fine-tuning improved model performance, enhancing diagnostic accuracy.
Area of Science:
- Neuroimaging
- Artificial Intelligence
- Medical Diagnostics
Background:
- Deep learning models demonstrate high accuracy in classifying amyloid brain scans for Alzheimer's disease (AD) diagnosis.
- Overfitting is a concern in deep learning models trained on limited datasets, potentially affecting generalization.
- Amyloid PET imaging is crucial for early AD diagnosis, aiding in treatment strategies.
Purpose of the Study:
- To develop and validate a [18F]Florbetaben amyloid brain PET scan classification model using a convolutional neural network (CNN).
- To assess the generalization potential of CNN models through external validation.
- To investigate methods for improving model performance when generalization is limited.
Main Methods:
- A CNN-based classification model was developed using the Dong-A University Hospital (DAUH) dataset.
- Three CNN architectures (Inception3D, ResNet3D, VGG3D) were evaluated.
- External validation was performed using the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset.
- Preprocessing included spatial normalization, count normalization, and skull stripping; smoothing was applied only to the external dataset.
- Models were fine-tuned using the external dataset to improve performance.
Main Results:
- Internal evaluation accuracy for Inception3D, ResNet3D, and VGG3D was 95.4%, 92.0%, and 97.7%, respectively.
- External validation accuracy was significantly lower: 76.7% for Inception3D, 67.1% for ResNet3D, and 85.3% for VGG3D.
- Fine-tuning with the external dataset improved Inception3D performance by 15.3% and ResNet3D by 16.9%.
Conclusions:
- CNN models show potential for [18F]Florbetaben amyloid brain PET scan classification.
- External validation is critical for assessing the generalization capability of deep learning models.
- Model structure modification and fine-tuning are effective strategies to enhance classification performance and address overfitting in neuroimaging AI.
