Intracranial aneurysms, bulges in cerebral blood vessels, pose a significant risk of rupture, leading to subarachnoid hemorrhage (SAH) with high mortality and morbidity. Early detection is crucial for timely intervention. However, manual detection is time-consuming, challenging, and prone to variability, particularly for small aneurysms. This study addresses this challenge by developing a deep learning-based approach for automated aneurysm detection and segmentation using 3D TOF-MRA. The use of deep learning offers the potential to improve diagnostic accuracy, reduce radiologist workload, and ultimately improve patient outcomes. Previous studies have explored 2D or 2D-projection based methods, but this study leverages the full 3D information in TOF-MRA for improved detection of 3D aneurysms. The objective is to develop and validate a robust CNN model for automated detection and segmentation of intracranial aneurysms, incorporating strategies to overcome the inherent data imbalance in clinical datasets.

Several studies have used conventional image processing techniques and deep learning for semi-automatic or automatic aneurysm detection. Conventional methods often rely on manual intervention and show limitations in accuracy and efficiency. CNNs, particularly U-Net architectures, have demonstrated success in medical image segmentation. Previous deep learning approaches for aneurysm detection have used 2D patches or 2D projections (MIPs) of TOF-MRA data, potentially limiting their ability to detect complex 3D aneurysms. This study builds upon these existing methods, focusing on a 3D approach and addressing the challenge of data imbalance inherent in aneurysm datasets where the number of aneurysms is significantly lower than normal vessel segments.

This retrospective study used 154 3D TOF-MRA datasets with intracranial aneurysms from Seoul National University Bundang Hospital (SNUBH), divided into training, validation, and testing sets. An external validation set of 113 TOF-MRA images from the ADAM challenge dataset was also used. Preprocessing steps included skull-stripping, signal intensity normalization, and N4 bias correction. 3D patches (64x64x64 voxels) were extracted along vessel skeletons. The ratio of normal to aneurysmal patches varied (1:1 to 1:5) to study the effect of data imbalance. A 3D U-Net with an auxiliary classifier was trained for semantic segmentation and auxiliary classification to handle data imbalance. The model was further refined by using residual connections and dual attention blocks within the U-Net architecture for improved performance. Data augmentation techniques (flips, zooming, noise addition, rotation, blurring, contrast, and gamma correction) were applied to increase training data size. The model's performance was evaluated using metrics such as accuracy, sensitivity, PPV, NPV, and DSC. The hyperparameters were tuned using the validation set, and the final model was tested using a separate test set.

Internal validation (SNUBH dataset) yielded an accuracy of 0.910 with a 2:1 ratio of normal to aneurysmal patches. The external validation (ADAM challenge dataset) achieved an accuracy of 0.883 with the same patch ratio. The auxiliary classifier significantly improved model accuracy compared to models without it across various patch ratios. The DSC was 0.701 ± 0.217 for the optimal patch ratio (2:1) in the internal dataset. Analysis by aneurysm size in the external dataset showed the highest accuracy (0.885) for aneurysms smaller than 5 mm. The model demonstrated good sensitivity and minimal false positives, suggesting its suitability for clinical use. Even for aneurysms smaller than 2 mm, an accuracy of approximately 80% was achieved. The study highlights the effective use of skeletonization to reduce the patch-level imbalance.

The results demonstrate the effectiveness of the proposed 3D patch-based CNN model for automated intracranial aneurysm detection and segmentation. The use of skeleton-based patch extraction and multi-task learning with an auxiliary classifier effectively addresses the data imbalance problem, leading to improved performance compared to traditional methods. The high accuracy and sensitivity, particularly for small aneurysms, make this model a promising tool for assisting radiologists in clinical practice. The model's good generalization ability, as demonstrated by its performance on the external dataset, enhances its potential for widespread clinical adoption. The study's findings contribute to the field of medical image analysis by showcasing the successful integration of 3D image analysis techniques with deep learning to improve the diagnosis of intracranial aneurysms.

This study successfully developed a deep learning model for automated intracranial aneurysm detection in 3D TOF-MRA. The novel approach, employing skeleton-based 3D patches, semantic segmentation, and auxiliary classification, effectively addresses data imbalance and improves accuracy. The model demonstrates good performance on both internal and external datasets, highlighting its potential for clinical use. Future research could focus on multi-center studies with larger, more diverse datasets, and exploration of advanced deep learning architectures.

The study is limited by the use of a single-center dataset for internal validation. Although external validation was performed using the ADAM challenge dataset, a multi-center study with diverse imaging protocols and scanner types is necessary to validate the model's generalizability further. The relatively small sample size, despite data augmentation, could also affect the model's robustness. Future work should focus on expanding the dataset and exploring different CNN architectures.