Automated Data Quality Control in FDOPA brain PET Imaging using Deep Learning.

Area of Science:

• Biomedical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Biomedical imaging research increasingly relies on large datasets, necessitating automated quality control (QC) methods.
• Existing AI-based QC approaches for neuroimaging (EEG, MRI) have limited replication across different domains.
• Automated QC is crucial for ensuring data integrity and reliability in multi-site and large-scale studies.

Purpose of the Study:

• To assess the feasibility of an automated QC pipeline for brain [18F]-FDOPA Positron Emission Tomography (PET) imaging.
• To evaluate the performance of Convolutional Neural Networks (CNNs) in identifying spatial misalignment and signal-to-noise ratio (SNR) issues in [18F]-FDOPA PET scans.
• To determine the potential of AI for operator-free QC in dopamine system biomarker imaging.

Main Methods:

• Two CNNs were developed and combined to analyze spatial misalignment and SNR in 200 manually QC'd [18F]-FDOPA PET scans.
• An additional 400 scans were created with simulated misalignment (200) and low SNR (200).
• A cross-validation approach (80% training, 20% validation) and out-of-sample validation with two independent datasets were employed.

Main Results:

• CNNs achieved high accuracy in the training dataset (motion: 0.86 ± 0.01, SNR: 0.69 ± 0.01).
• The automated QC pipeline demonstrated 100% accuracy when applied to two independent out-of-sample datasets.
• Data dimensionality reduction negatively impacted CNN generalizability, particularly when transitioning from 3D to 1D datasets.

Conclusions:

• Automated quality control of [18F]-FDOPA PET imaging using CNNs is feasible.
• This AI-driven approach shows potential for extension to other PET tracers and applications (brain and non-brain).
• Successful implementation is contingent upon the availability of large, diverse datasets for robust algorithm training.