Multimodal Deep Learning and Knowledge-Enhanced Intelligent Decision Support System for Pipeline Embolization Device Size Selection in Intracranial Aneurysm Treatment.
This study aimed to develop an end-to-end intelligent decision support system, NeurAneuNet, to automate the selection of Pipeline Embolization Device (PED) size and landing zones for intracranial aneurysm treatment, thereby reducing reliance on operator experience and improving planning consistency.
NeurAneuNet integrates multimodal deep learning with knowledge enhancement. The system processes three-dimensional rotational angiography (3DRA) images using a dual-path attention U-Net++ for aneurysm segmentation, extracts knowledge-augmented geometric and clinical features, fuses five feature modalities via tensor decomposition, and predicts optimal PED sizing and landing zones with a high-order Kolmogorov-Arnold Network (KAN). A dataset of 600 aneurysms (including 210 PED-treated cases) was used for model development and validation, with an independent clinical cohort of 21 cases employed to assess real-world utility.
NeurAneuNet achieved a Dice coefficient of 0.874 ± 0.03 for aneurysm segmentation, a PED size classification accuracy of 91.8%, and a diameter prediction error of 0.24 ± 0.10 mm. In the independent test, its primary recommendation agreed with the expert consensus in 95.2% of cases. When assisting clinicians, the system reduced planning time by 44.8% and significantly lowered subjective cognitive workload, as assessed by the National Aeronautics and Space Administration Task Load Index (NASA-TLX) score (from 33 ± 8 to 21 ± 5).
NeurAneuNet demonstrates clinically relevant accuracy and efficiency in automating PED treatment planning, providing robust, intelligent support for intracranial aneurysm interventions. The system provides a promising foundation for developing a generalizable AI-driven decision support framework in complex medical scenarios.
NeurAneuNet integrates multimodal deep learning with knowledge enhancement. The system processes three-dimensional rotational angiography (3DRA) images using a dual-path attention U-Net++ for aneurysm segmentation, extracts knowledge-augmented geometric and clinical features, fuses five feature modalities via tensor decomposition, and predicts optimal PED sizing and landing zones with a high-order Kolmogorov-Arnold Network (KAN). A dataset of 600 aneurysms (including 210 PED-treated cases) was used for model development and validation, with an independent clinical cohort of 21 cases employed to assess real-world utility.
NeurAneuNet achieved a Dice coefficient of 0.874 ± 0.03 for aneurysm segmentation, a PED size classification accuracy of 91.8%, and a diameter prediction error of 0.24 ± 0.10 mm. In the independent test, its primary recommendation agreed with the expert consensus in 95.2% of cases. When assisting clinicians, the system reduced planning time by 44.8% and significantly lowered subjective cognitive workload, as assessed by the National Aeronautics and Space Administration Task Load Index (NASA-TLX) score (from 33 ± 8 to 21 ± 5).
NeurAneuNet demonstrates clinically relevant accuracy and efficiency in automating PED treatment planning, providing robust, intelligent support for intracranial aneurysm interventions. The system provides a promising foundation for developing a generalizable AI-driven decision support framework in complex medical scenarios.
Authors
Wen Wen, Guo Guo, Peng Peng, Chen Chen, Huangfu Huangfu, Zhao Zhao, Gao Gao, Ni Ni, Zhang Zhang, Liu Liu, Liu Liu, Liang Liang
View on Pubmed