Synthetic vascular structure generation for unsupervised pre-training in CTA segmentation tasks
This addresses the problem of limited labeled data for medical imaging segmentation, particularly for stroke treatment, but is incremental as it builds on existing unsupervised pre-training methods.
The researchers tackled the scarcity of labeled CT data for vessel segmentation by generating synthetic vascular structures to pre-train models, achieving an increase in accuracy compared to models trained only on hand-labeled data.
Large enough computed tomography (CT) data sets to train supervised deep models are often hard to come by. One contributing issue is the amount of manual labor that goes into creating ground truth labels, specially for volumetric data. In this research, we train a U-net architecture at a vessel segmentation task that can be used to provide insights when treating stroke patients. We create a computational model that generates synthetic vascular structures which can be blended into unlabeled CT scans of the head. This unsupervised approached to labelling is used to pre-train deep segmentation models, which are later fine-tuned on real examples to achieve an increase in accuracy compared to models trained exclusively on a hand-labeled data set.