Deep-learning-based cardiac amyloidosis classification from early acquired pet images

Area of Science:

• Nuclear Medicine
• Artificial Intelligence
• Cardiology

Background:

• Cardiac amyloidosis (CA) diagnosis can be challenging.
• Early detection is crucial for effective management.
• Novel imaging analysis techniques are needed.

Purpose of the Study:

• To evaluate deep learning for characterizing cardiac amyloidosis.
• To assess the potential of early [18F]-Florbetaben PET images.
• To develop a deep convolutional neural network (CAclassNet) for CA diagnosis.

Main Methods:

• Utilized early (15 min post-injection) [18F]-Florbetaben PET/CT images from 47 subjects (13 ATTR-CA, 15 AL-CA, 19 controls).
• Developed CAclassNet, a deep convolutional neural network, to classify images into AL-CA, ATTR-CA, and control groups.
• Trained, validated, and tested the network on 1107 2D images.

Main Results:

• CAclassNet achieved high cross-validation accuracy (95.49%) and good test set performance.
• Sensitivity, specificity, and accuracy for AL-CA were 1, 0.912, and 0.936, respectively.
• Sensitivity, specificity, and accuracy for ATTR-CA were 0.935, 0.897, and 0.972, respectively.

Conclusions:

• Deep learning models like CAclassNet show significant promise for diagnosing cardiac amyloidosis.
• Early [18F]-Florbetaben PET imaging can be effectively analyzed by AI for CA detection.
• CAclassNet can serve as a valuable clinical aid for CA diagnosis.