AUTOMATED INTRACRANIAL ANEURYSM SEGMENTATION IN TIME-OFFLIGHT MR ANGIOGRAPHY USING A FINE-TUNED SEGMENT ANYTHING MODEL: COMPARISON WITH NON-RADIOLOGIST PHYSICIANS

PURPOSE: This study aimed to evaluate the performance of a Segment Anything Model (SAM)-based Artificial Intelligence (AI) model fine-tuned for automatic detection and segmentation of intracranial aneurysms on Time-of-Flight MR angiography (TOF-MRA).
MATERIALS AND METHOD: The AI model, fine-tuned on 397 TOF-MRA scans, enhances SAM by integrating a convolutional neural network-based encoder and a self-prompting module for tiny lesion segmentation. The AI model and two non-radiologist physicians assessed 30 aneurysmcontaining TOF-MRA scans not used in training. We compared patient- and lesion-wise sensitivity, lesion-wise positive predictive value (PPV), segmentation accuracy (Dice Similarity Coefficient, DSC), and patient-wise segmentation time. DSC was measured based on manual annotations by a neuroradiologist with 15 years of experience. The same comparisons were conducted in subgroups stratified by aneurysm size (small <5 mm vs. large ≥5 mm), location, and lobularity.
RESULTS: Among the 30 TOF-MRA scans, a total of 52 aneurysms were detected and segmented. Of these, 84.6% were small, 40.4% were in the middle cerebral artery, and 78.8% were unilobular. The AI model achieved 76.7% patient-wise sensitivity, 82.7% lesion-wise sensitivity, 84.3% PPV, and a median DSC of 69.7 (IQR 48.4-78.4), showing comparable performance to human readers—Reader 1: 63.3%, 69.2%, 80.0%, and 71.2 (IQR 52.3-81.7), respectively; Reader 2: 63.3%, 69.2%, 85.7%, and 73.4 (IQR 62.3-77.8), respectively. The AI model completed segmentation in a median time of 0.74 minutes (IQR 0.65-0.79), notably faster than human readers. In size-stratified analysis, the AI model showed a trend toward higher sensitivity (81.8%) for small aneurysms than human readers (both 63.6%). The AI model yielded similar segmentation accuracy between small and large aneurysms, while human readers tended to show higher accuracy for large aneurysms. No significant differences were found within each location or lobularity subgroup.
CONCLUSION: The SAM-based AI model showed a tendency toward higher sensitivity for detecting small aneurysms than non-radiologist physicians. It also achieved consistent segmentation performance across aneurysm sizes, unlike human readers whose accuracy showed a declining trend for small aneurysms. Additionally, it completed segmentation markedly faster than the physicians. Fine-tuned SAM-based AI model may assist non-expert clinicians in detecting small aneurysms reliably, particularly in settings with limited neuroradiologist support.