A Mask R-CNN-Based Approach for Brain Aneurysm Detection and Segmentation from TOF-MRA Data

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Accurate detection of intracranial aneurysms, particularly small ones (<3 mm), is challenging in Time-of-Flight Magnetic Resonance Angiography (TOF-MRA) due to subtle features and low contrast.
• Existing methods struggle with sensitivity for micro-aneurysms, necessitating advanced detection techniques.

Purpose of the Study:

• To develop and evaluate a Mask R-CNN-based deep learning framework for automated detection and segmentation of intracranial aneurysms.
• To enhance the model's sensitivity for small aneurysms through specific architectural modifications.

Main Methods:

• Utilized a dataset of 447 TOF-MRA volumes (161 aneurysmal, 286 healthy) with 5-fold cross-validation.
• Implemented Bayesian hyperparameter optimization (Optuna) and introduced a Small Object Aware ROI Head and customized anchors.
• Incorporated healthy scans as negative samples and applied targeted data augmentation for improved generalization.

Main Results:

• Achieved a Dice coefficient of 0.8832, precision of 0.9404, and sensitivity of 0.8677.
• Demonstrated consistent performance across various aneurysm sizes, including micro-aneurysms.
• The model proved effective in enhancing background modeling and region proposal quality.

Conclusions:

• The developed deep learning tool offers a clinically viable solution for intracranial aneurysm detection.
• Architectural innovations and automated optimization contribute to a reliable system for assisting radiologists.
• The framework shows potential as a second-reader system, improving diagnostic accuracy in neurovascular imaging.