Automated detection of intracranial aneurysms using skeleton-based 3D patches, semantic segmentation, and auxiliary classification for overcoming data imbalance in brain TOF-MRA

Cerebral aneurysms are bulges of cerebral blood vessels that can leak or rupture, causing subarachnoid hemorrhage (SAH) with high mortality and morbidity. Approximately 3% of healthy adults harbor an intracranial aneurysm, and rupture risk is associated with factors such as genetics, sex, age, and aneurysm site, size, and shape. Early and accurate automated detection could reduce diagnostic variability and workload for radiologists, especially for small aneurysms that are difficult to detect. Prior automated approaches have often relied on 2D images or maximum intensity projections, which may limit performance on inherently 3D aneurysm morphology. Therefore, the study aimed to develop and validate a 3D CNN-based method using TOF-MRA that detects and segments intracranial aneurysms from skeleton-based 3D patches, employing semantic segmentation with an auxiliary classifier to address severe data imbalance, and to preliminarily assess clinical usefulness.

Conventional image processing approaches for semi-automatic aneurysm detection have been proposed, but deep learning has transformed medical imaging tasks. Prior works include 2D patch-based ResNet detection on TOF-MRA (Ueda et al.), U-Net trained on MIP images to predict aneurysm size (Stember et al.), and CAD systems on MIPs with CNN classifiers (Nakao et al.). Other studies demonstrated deep learning for aneurysm detection/segmentation on TOF-MRA and CTA, and the ADAM challenge compared methods for unruptured aneurysm detection/segmentation on TOF-MRA. However, many existing methods rely on 2D or MIP representations, potentially limiting 3D aneurysm detection capability, motivating a fully 3D approach.

Study design and datasets: Retrospective single-center study approved by the SNUBH IRB with consent waived. Internal dataset: 154 TOF-MRA volumes with intracranial aneurysms (contrast-enhanced angiographic data from Siemens Axiom Artis and GE Innova IGS 630). Acquisition parameters: slice thickness 0.2–0.5 mm, matrix 1024×1024, voxel size 0.2–0.7 mm. Data split: 120 train, 19 validation, 15 test. Demographics: age 32–76 years (mean 53.90 ± 12.97), 70% female. Diagnoses: mostly unruptured aneurysms (n=150), one non-aneurysmal anatomical variant, three SAH due to Pcom aneurysms. Aneurysm sizes 1.8–32.6 mm (mean 2.6 ± 1.9 mm); 130 <5 mm, 21 between 5–10 mm, 3 >10 mm. Ground truth segmentation masks were semi-automatically delineated on 3D contrast-enhanced T1-weighted images by a neuroradiologist (>10 years) using MITK, and validated by another neuroradiologist (>18 years). External dataset: 113 TOF-MRA cases from the ADAM challenge (UMC Utrecht, Philips 1–3T, in-plane voxel spacing 0.195–1.04 mm, slice thickness 0.4–0.7 mm, 2001–2019), including subjects with and without aneurysms. Aneurysm annotations were created slice-wise from neck to dome by an interventional neuroradiologist (>10 years), converted to binary masks. Pre-processing: Intensity clipping to [0.5, 99.5] percentiles; z-score normalization per volume; N4 bias field correction; skull stripping using a trained deep-learning BET model. Resampling to 0.3×0.3×0.3 mm isotropic voxels. Vessel skeletonization through thresholding to binary, connected component labeling (~3000-pixel seed structures), region growing, morphological closing, and iterative 3D thinning until skeletal convergence. Patch extraction: 3D patches of size 64×64×64 extracted along the vessel skeleton. Positive patches were randomly sampled non-centered around the aneurysm ensuring full inclusion of manual mask. Negative and positive patch ratios varied from 1:1 to 1:5 to study class imbalance. Data augmentation included flips, zoom, rotation, blur, contrast/gamma adjustments, and Rician/Gaussian noise. Model architecture: Base on 3D U-Net with residual basic blocks and a dual attention block to focus on informative regions/features. An auxiliary classifier was added along the decoder network (bridge block) to mitigate gradient loss and improve convergence, using binary cross-entropy loss; segmentation used Dice and Tversky focal loss plus cross-entropy, with final loss as an average of classifier and segmentation losses at inference/training. Comparisons were made to a standard 3D U-Net and nnU-Net configurations. Training details: Batch size 16; batch normalization; parametric ReLU activations; 500 epochs; optimizer Adam (lr=0.001). Training sets were constructed with varying normal:aneurysm patch ratios (1:1 to 5:1). Validation used for hyperparameter tuning; test sets were held out. Evaluation metrics: Accuracy, sensitivity (recall), positive predictive value (PPV), negative predictive value (NPV), and Dice similarity coefficient (DSC) computed at patient/segmentation levels to assess detection and segmentation performance. Statistical comparisons used t-tests with significance at p<0.05.

• Internal validation (SNUBH): Best performance at normal:abnormal patch ratio of 2:1 with accuracy 0.910, specificity 0.893, PPV 0.896, NPV 0.909, sensitivity 0.926, DSC 0.701 ± 0.217. Adding the auxiliary classifier improved accuracy over models without it. DSC increased as the ratio rose from 1:1, then decreased at 5:1.
• External validation (ADAM): Highest accuracy achieved with auxiliary loss and a 2:1 normal:abnormal patch ratio; reported external accuracy 0.883. By aneurysm size, mean segmentation accuracy for <5 mm aneurysms reached 0.885.
• Overall per-patient evaluation (discussion summary): Mean DSC 0.755 ± 0.09, sensitivity 0.882, and false positives (FP) 0.305, indicating good sensitivity with low FP burden.
• The model effectively detected and segmented very small aneurysms (<10 mm), with approximately 80% accuracy even for aneurysms smaller than 2 mm.
• Multi-task learning with an auxiliary classifier and skeleton-based patching addressed severe class imbalance and reduced false positives while maintaining segmentation performance.

The study demonstrates that a 3D CNN trained on skeleton-based patches with semantic segmentation and an auxiliary classification head can reliably detect and segment intracranial aneurysms in TOF-MRA, addressing the clinical challenge of small aneurysm detection and reader variability. The combination of vessel skeletonization (reducing search dimensionality and patch imbalance) and multi-task learning (segmentation plus classification) improved learning stability under severe class imbalance, outperforming a standard 3D U-Net on the same data. The model achieved high accuracy and sensitivity internally and externally, including strong performance on small aneurysms (<5 mm), suggesting potential utility in reducing radiologist workload and diagnostic fatigue. Comparative context indicates advances over prior 2D/MIP-based methods by leveraging full 3D information. Failure analyses revealed false positives and false negatives arising from aneurysms much smaller than average training sizes and from low lesion conspicuity (subtle contrast/brightness), highlighting areas for further robustness improvements.

A 3D patch-level multi-task learning framework—semantic segmentation with an auxiliary classifier—was developed for automatic detection and segmentation of intracranial aneurysms on TOF-MRA. Using vessel skeleton-based patch extraction, comprehensive pre-processing, and enhanced 3D U-Net architecture, the method achieved high accuracy and sensitivity with a low false-positive rate in both internal and external validations, including reliable performance on very small aneurysms. The approach shows promise for rapid, reliable clinical screening and delineation of aneurysms along brain vessels. Future work should include broader multi-center validation and exploration of advanced 3D encoder–decoder variants to further improve generalizability and performance.

• Single-center internal dataset limits generalizability; multi-center studies across different institutions, scanners, and protocols are needed.
• Relatively small sample size; although extensive augmentation was used, more patient data would strengthen training robustness.
• Sensitivity to very small lesions and low-contrast cases can cause false negatives/positives; further architectural and training refinements may be required.
• Only specific architectures were explored; additional 3D encoding–decoding variants could provide further gains.