[공저자] Lee et al., 2021, Performance evaluation in [18F]Florbetaben brain PET images classification using 3D Convolutional Neural Network

# 세줄 요약 #

1. 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 (ADNI) dataset.
2. Three types of models were used, depending on their structure: The accuracy of model evaluation for Inception 3D, ResNet 3D, VGG 3D was 95.4%, 92.0%, 97.7%, and the accuracy of the external validation was 76.7%, 67.1%, 85.3%, respectively.
3. When external validation was performed again after fine-tuning (retrain), the performance improved to 15.3% for Inception 3D and 16.9% for ResNet 3D, which means changing the structure or fine-tuning the model can help improve the classification performance.

 

# 상세 리뷰 #

1. Introduction

• As there is no effective treatment for Alzheimer's Disease (AD), an accurate diagnosis is essential for developing the patient's future treatment plan.
  • Representative radiotracers used to diagnose AD is PET
    • [18F]fluro-D-glucose: identifying the degree of brain metabolism.
    • [18F]Florbetaben, [18F]Florbetapir, [18F]Fluntemetamol: observing brain amyloid plaque load.
• Previous Study, an AD diagnosis classifier.
  • Using PCA and SVM was utilized after image dimension reduction of [18F]Florbetaben: Cho et al., 2019
  • Using Visual Geometry Group (VGG) CNN model to amyloid deposition classification (89% Accuracy): Kang et al., 2018 
• Because of the nature of medical images, acquisition costs can be high.
  • It is not easy to construct a large dataset.
  • To confirm the generalization possibility, it is necessary to configure an external dataset.
• In this study,
  • Using brain PET scans of [18F]Florbetaben
    • radiotracer that visualizes the classification of B-amyloid (A_beta)
    • A_beta: the main component of amyloid plaques found in the brain.
  • Using 3D voxel input CNN models:
    • Inception, VGG, ResNet.
  • Dataset composition 
    • Training & in-house validation: acquired from Dong-A University Hospital (DAUH)
    • External validation: acquired from Alzheimer's Disease Neuroimaging Initiative (ADNI)

 

2. Materials and Methods

2.1. Data acquisition

• The DAUH dataset
  • 432 subjects of [18F]Florbetaben PET
    • 40mCT Flow PET/CT Scanner, 3mm FWHM Gaussian filter
  • Label: A_beta negative & A_beta positive
    • The decision of a nuclear medicine specialist at DAUH
• The ADNI (Alzheimer's Neuroimaging Initiative) dataset
  • External dataset for these study
  • These scans were preprocessed as an ADNI internal protocol.
    • co-registration, averaging, size changing, standardization, smoothing
    • To achieve the same condition, 3mm FWHM Gaussian filter (using statistical parametric mapping (SPM))
  • Label: A_beta negative & A_beta positive
    • The decision of a nuclear medicine specialist at DAUH

 

2.2. Image processing in common

• Step 1. Spatial Normalization: registration of the original image to a specific PET template
  • A reference brain PET template: Co-registration was performed with a PET image and a paired CT image obtained from a subject.
  • After image registration, the tissue is stretched or compressed to fit the template brain.
• Step 2. Count Normalization:
  • normalizes the entire observation area to the value of the area representing non-specific, lesion-independent absorption,
  • allowing absolute and relative comparisons in specific absorption areas of the patient-patient image
• Step 3. Skull stripping:
  • non-brain tissues are considered an obstacle in brain image analysis
  • skull stripping is required for brain imaging analysis study.

 

2.3. Model architecture

• We apply 3D CNNs of three well-known architectures to the amyloid classification
  • Inception, ResNet, VGG19

 

2.4. Model selection and evaluation

• In DAUH dataset: 80% Training, 20% model evaluation
  • 4-fold cross-validation with the training dataset (192 positive & 153 negative)
  • we could select the lowest loss model out of the four performances.
• In the external dataset: using the same sample extraction method.

 

3. Results

3.2. Model evaluation

• The model evaluation results of the Inception3D, ResNet3D, VGG3D models after training the models with the DAUH dataset.

 

3.3. External validation

• With respect to accuracy and AUC -> VGG3D showed the best classification performance.

 

3.4. Retraining model.

• Redefining the model to create a generalized model by additionally retraining part of the external dataset.
  • using 38 positive & 50 negatives in ADNI dataset(external dataset) for fine-tuning.
  • the classification performance was improved compared to before retraining.

 

* Reference: Lee S-Y, Kang H, Jeong J-H, Kang D-y (2021) Performance evaluation in [18F]Florbetaben
brain PET images classification using 3D Convolutional Neural Network. PLoS ONE 16(10):
e0258214. https://doi.org/10.1371/journal.pone.0258214