Synthetic Vascular Image Selection For Deep Learning Based Cerebral Bifurcation Classification

Deep learning algorithms rely heavily on large datasets to efficiently perform various pattern recognition tasks. However, collecting ground truth datasets, which include the necessary annotations for training neural networks, is often a challenging and labor-intensive process. Sometimes, in order to alleviate the labeling burden, while still providing high quality augmented data, synthetic models are being used. A properly designed synthetic model can prove very efficient for various pattern recognition tasks, subject to a thorough mimicking of the actual ground truth. However, when exploiting synthetic images, one might encounter significant drawbacks: how can we ensure that the synthetic data contributes positively to the neural network training process ? Are the synthetic images of sufficient quality to be useful ? Or should some be discarded from the training dataset (as they may lessen the CNN learning ability) ? In this work, we propose to run a subjective experiment to assess the similarities between vascular bifurcations, we can hence sort various bifurcations, may they be ground truth portions of the vascular tree as acquired on Magnetic Resonance Angiography (MRA) - Time of Flight (ToF) acquisitions, or synthetically modeled bifurcations. Once the subjective experiment was set up and conducted, we tested various objective quality measures to assess the fidelity of the synthetic images. More precisely, these automatic quality estimation metrics were used to remove any malfunctioning synthetic model from the training dataset. A CNN is finally trained on a bifurcation classification task with either the full training set or its reduced version (after quality filtering).