Background and Purpose: Rapid and accurate identification of vessel occlusion is critical for timely intervention and treatment in ischemic stroke patients. Manual assessments of Computed Tomography Angiograms (CTAs) by radiologists is time consuming, subject to inter-observer variability, and challenging in emergency settings. This work presents a deep-learning approach based on nnUNet, a self-configuring convolutional neural network (CNN), trained on Maximum Intensity Projection (MIP) volumes derived from CTA volumes to detect and segment large as well as distal vessel occlusion. Methodology: We curated 500 de-identified brain CTA studies from a diverse multi-institutional dataset based on radiology reports, comprising 450 occlusion-positive and 50 occlusion-negative cases. Only axial series with slice thickness < 0.7 mm were included. Preprocessing steps performed on these CTA volumes included skull stripping, vessel enhancement, creation of Maximum Intensity Projection (MIP) and using histogram based windowing technique to enhance vessels in the MIP volumes. Manual annotation of the occlusion sites was performed and validated by a neurointerventional radiologist. The resulting MIP volumes, along with the corresponding annotated masks are used as input for training the deep learning model. A 3D convolutional neural network deep learning model based on nnU-Net is trained to perform voxel-wise segmentation of large vessel occlusions using patch-based processing. The training optimization and architectural refinement are currently ongoing, and full performance evaluation will be conducted upon completion. Results: The model was evaluated for both classification and localization accuracy on the test dataset. The model achieved perfect 100% classification accuracy for both no occlusion and large vessel occlusion (LVO) cases, and 88% accuracy for finer vessel occlusions. Segmentation performance was similarly high, with 100% accuracy for LVOs and 82% for finer vessels. Efforts are currently underway to refine the training strategy to further improve accuracy, particularly for distal vessel occlusions. Conclusion: Our nnU-Net–based model enables rapid and precise localization of vessel occlusions in brain CTA, facilitating timely stroke triage and treatment planning. The MIP-driven pipeline demonstrates high efficiency and accuracy, with strong potential for seamless integration into PACS workflows in emergency care settings.
Yousouf et al. (Thu,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: