Introduction 3D visualization of cerebral arteries allows better detection and analysis of neurovascular diseases. While deep learning (DL) models enable automatic segmentation of cerebral arteries on CT, MRI and 3DRA independently, they have mostly been developed to process one imaging modality. This may limit the reproducibility and comparability of arteries segmented on different imaging modalities that a patient may undergo during follow-up.
Aim of Study To evaluate whether a unified ensemble DL model, trained on CT, MRI, and 3DRA, improves (i) the segmentation reproducibility between imaging modalities and (ii) the segmentation quality for each modality.
Methods We developed an ensemble of DL models to segment cerebral arteries on CT, MRI, and 3DRA independent of imaging modality. We trained this model on a large dataset of CT, MRI, and 3DRA whose arteries had been manually segmented by a neuroradiologist. The model was prospectively evaluated on a dataset of 50 patients with matched CT, MRI and 3DRA. We compared the segmentation quality of this unified ensemble model with models trained to segment only one image modality (CT, MRI or 3DRA).
Results The unified ensemble DL model improved segmentation reproducibility between the different imaging modalities compared to models trained for on a single image type. It also offers finer segmentation of cerebral arteries on CT compared to a simpler model trained only on that modality.
Conclusion A unified ensemble DL model allows for better quality and reproducibility of cerebral arteries segmentation on different imaging modalities, which may improve comparability in follow-up imaging.
Disclosure of Interest Nothing to disclose
Statistics from Altmetric.com
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.