Synthetic Vascular Image Selection For Deep Learning Based Cerebral Bifurcation Classification.

Area of Science:

• Medical imaging analysis
• Deep learning in healthcare
• Computational anatomy

Background:

• Deep learning requires large, annotated datasets, which are costly and time-consuming to acquire.
• Synthetic data generation offers a solution to augment datasets but raises concerns about data quality and its impact on model training.
• Ensuring the fidelity of synthetic medical images is critical for their effective use in neural network training.

Purpose of the Study:

• To assess the similarity between real and synthetic vascular bifurcations using subjective evaluation.
• To develop and apply objective quality measures for filtering synthetic medical images.
• To evaluate the impact of filtered synthetic data on the performance of a convolutional neural network (CNN) for bifurcation classification.

Main Methods:

• Conducted a subjective experiment to compare real (MRA-ToF) and synthetic vascular bifurcations.
• Applied objective quality assessment metrics to evaluate the fidelity of synthetic images.
• Utilized automatic quality estimation metrics to filter out low-quality synthetic data.
• Trained a CNN on both the full and filtered synthetic datasets for bifurcation classification.

Main Results:

• Subjective experiments provided a basis for sorting and evaluating bifurcations.
• Objective quality measures successfully identified and enabled the removal of malfunctioning synthetic models.
• CNNs trained on the filtered dataset showed improved performance compared to those trained on the full dataset, indicating the benefit of quality control.

Conclusions:

• Subjective and objective assessments are valuable for ensuring the quality of synthetic medical data.
• Filtering synthetic data based on quality metrics can significantly enhance the performance of deep learning models.
• This approach addresses the challenge of integrating synthetic data into medical imaging AI pipelines effectively.