Accurate identification and segmentation of intracranial aneurysms in brain CTA scans remain challenging due to their small size, low contrast, and high morphological variability. This paper proposes a label-efficient deep learning framework that integrates synthetic mask generation, vessel-enhancement preprocessing, and a modified Attention U-Net augmented with Convolutional Block Attention Modules (CBAM). To address class imbalance and enhance sensitivity to small aneurysm regions, the network is optimized using Tversky loss. In the absence of manual annotations, automatically generated vessel masks are used as synthetic supervisory signals, enabling training under limited annotation conditions. Experimental results demonstrate that the proposed model outperforms a standard U-Net, achieving Dice = 0.8798, IoU = 0.8056, Precision = 0.8327, Recall = 0.9152, and F1-score = 0.8584, while producing smoother and more continuous vessel segmentations with reduced background noise. These findings suggest that combining spatial-channel attention mechanisms with false-negative-aware optimization can improve aneurysm segmentation performance and provide a reproducible, annotation-efficient framework for future neurovascular imaging applications.