Abstract
A paucity of convolutional neural network (CNN) studies of workflow analysis exists in microneurosurgery for middle cerebral artery (MCA) aneurysm clipping, underscoring the need to develop models to identify procedural steps for educational and analytical purposes.
Forty-two MCA aneurysm clipping videos from 1 neurosurgeon (October 2020-October 2024) were retrospectively collected; 25 videos (38 517 frames) with uncomplicated MCA aneurysms were selected for evaluation. Twelve surgical steps were identified: sylvian fissure splitting, M4 segment to M2 branch dissection, M1 segment preparation, supraclinoid internal carotid artery dissection, A1 anterior cerebral artery dissection, aneurysm neck dissection, aneurysm mobilization, proximal control, clip application, occlusion confirmation, release of proximal control, and aneurysm resection. Indocyanine green videoangiography was its own category. Two neurosurgeons annotated and extracted 1 frame/second, resized to 224 × 224 pixels; a CNN long short-term memory model was constructed to capture spatial and temporal features. Eighteen operative cases (27 713 frames) were used for training and 7 (10 804 frames) for testing. Top-1 macro averages were determined to assess performance, including accuracy, precision, recall, and F1 score.
The model achieved high accuracy for categories with extended durations or distinctive visual cues, including sylvian fissure splitting (precision, 77.6%; recall, 89.9%; F1 score, 83.3%) and indocyanine green videoangiography (precision, 98.7%; recall, 97.2%; F1 score, 97.9%). Steps that occurred briefly or infrequently, such as M1 segment preparation (precision, 100%; recall, 5.5%; F1 score, 10.4%) and occlusion confirmation (precision, 75.0%; recall, 1.2%; F1 score, 2.3%), had lower accuracy. The macro-averaged results were precision, 79.8%; recall, 36.9%; and F1 score, 44.0%.
The CNN long short-term memory approach to identify steps in MCA aneurysm clipping from microneurosurgery videos is feasible. Although certain phases were challenging, targeted data collection and refined annotation enhanced model performance. This work shows the feasibility of automated surgical-phase recognition in microneurosurgery; applications include resident training and operative video analysis.