Artificial intelligence-based automatic nidus segmentation of cerebral arteriovenous malformation on time-of-flight magnetic resonance angiography

Objective: Accurate nidus segmentation and quantification have long been challenging but important tasks in the clinical management of Cerebral Arteriovenous Malformation (CAVM). However, there are still dilemmas in nidus segmentation, such as difficulty defining the demarcation of the nidus, observer-dependent variation and time consumption. The aim of this study is to develop an artificial intelligence model to automatically segment the nidus on Time-Of-Flight Magnetic Resonance Angiography (TOF-MRA) images. Methods: A total of 92 patients with CAVM who underwent both TOF-MRA and DSA examinations were enrolled. Two neurosurgeons manually segmented the nidus on TOF-MRA images, which were regarded as the groundtruth reference. A U-Net-based AI model was created for automatic nidus detection and segmentation on TOFMRA images. Results: The mean nidus volumes of the AI segmentation model and the ground truth were 5.427 +/- 4.996 and 4.824 +/- 4.567 mL, respectively. The mean difference in the nidus volume between the two groups was 0.603 +/- 1.514 mL, which was not statistically significant (P = 0.693). The DSC, precision and recall of the test set were 0.754 +/- 0.074, 0.713 +/- 0.102 and 0.816 +/- 0.098, respectively. The linear correlation coefficient of the nidus volume between these two groups was 0.988, p < 0.001. Conclusion: The performance of the AI segmentation model is moderate consistent with that of manual segmentation. This AI model has great potential in clinical settings, such as preoperative planning, treatment efficacy evaluation, risk stratification and follow-up.