Cerebral AVM segmentation from 3D rotational angiography images by convolutional neural networks

• DOI: 10.1016/j.neuri.2023.100138
• Publication Date: 2023-07-22T07:33:14Z
• Authors: Mounir Lahlouh; Raphaël Blanc; Michel Piotin; Jérôme Szewczyk; Nicolas Passat; Yasmina Chenoune
• Keywords: [INFO.INFO-AI] Computer Science [cs]/Artificial Intelligence [cs.AI]; 3D rotational angiography; Focal Tversky; focal Tversky; [INFO.INFO-IM] Computer Science [cs]/Medical Imaging; 610; Cerebrovascular segmentation; Fully automatic segmentation; fully automatic segmentation; cerebrovascular segmentation; Neurosciences. Biological psychiatry. Neuropsychiatry; [INFO.INFO-AI]Computer Science [cs]/Artificial Intelligence [cs.AI]; [INFO.INFO-TI] Computer Science [cs]/Image Processing [eess.IV]; [INFO.INFO-TI]Computer Science [cs]/Image Processing [eess.IV]; 616; convolutional neural networks; [INFO.INFO-IM]Computer Science [cs]/Medical Imaging; Convolutional neural networks; RC321-571

3D rotational angiography (3DRA) provides high quality images of the cerebral arteriovenous malformation (AVM) nidus that can be reconstructed in 3D. However, these reconstructions are limited to only visualization without possible interactive exploration geometric characteristics structures. Refined understanding AVM angioarchitecture prior treatment is mandatory and vascular segmentation an important preliminary step allow physicians analyze complex networks help guide microcatheters navigation embolization AVM. A deep learning method was developed for 3DRA patients. The uses a fully convolutional neural network with U-Net-like architecture DenseNet backbone. compound loss function, combining Cross Entropy Focal Tversky, employed robust segmentation. Binary masks automatically generated from region-growing have been used train validate our model. able achieve vessels significantly outperformed algorithm. Our experiments were performed on 9 trained achieved Dice Similarity Coefficient (DSC) 80.43%, surpassing other U-Net like architectures algorithm manually approved test set by physicians. This work demonstrates potential learning-based characterizing very tiny structures even when training phase results automatic or semi-automatic method. proposed contribute planning guidance endovascular procedures.