A geometric-to-neural cascade for cerebral microbleed detection in susceptibility-weighted MRI

Technical Summary: A Geometric-to-Neural Cascade for Cerebral Microbleed Detection

Problem Statement
Cerebral microbleeds (CMBs) are critical imaging biomarkers for cerebral small vessel disease and cerebral amyloid angiopathy (CAA). However, their automated detection in susceptibility-weighted imaging (SWI) remains challenging. Existing methods struggle with high false-positive (FP) rates caused by vessel cross-sections, iron deposits, and calcifications, which share the hypointense appearance of CMBs. Furthermore, threshold-based segmentation is often unreliable due to variability in scanner types and acquisition protocols. A significant barrier to training robust deep learning models is the scarcity of publicly available, densely labeled SWI datasets that represent the high lesion burdens typical of CAA; most public datasets contain few CMBs per scan or rely on lower-resolution T2* GRE sequences. Additionally, obtaining ground-truth voxel-wise segmentation masks for training is time-consuming and requires specialist neuroradiologists.

Methodology
The authors propose a fully automated, three-stage cascade pipeline designed to operate with minimal human-in-the-loop supervision (binary "yes/no" labels per candidate rather than dense masks). The pipeline consists of:

1. Subject-Adaptive Unsupervised Candidate Generation:

   • Anatomical Masking: The system excludes physiologically irrelevant regions (ventricles, CSF, choroid plexus, cerebellum) and image edges using SynthSeg parcellations and a 5-voxel erosion of the brain mask.
   • Gaussian Mixture Model (GMM) Thresholding: A 3-component GMM is fitted to each subject's SWI intensity histogram to distinguish background, gray matter (GM), and white matter (WM). A subject-specific low-intensity threshold is defined at one standard deviation below the GM mean, allowing the system to adapt to varying contrast levels across different scanners.
   • Geometric Filtering: Candidates are generated via diagonal-constrained priority region growing. They are filtered by size (volume > 2 voxels), sphericity (bounding box diagonal < 10 mm), and aspect ratio (< 3.0) to eliminate elongated structures like vessels.

2. Stage A: High-Sensitivity Classification:

   • A lightweight 3D ResNet (661k parameters) classifies all geometric candidates as CMB or non-CMB.
   • The model is trained on 16x16x16 voxel patches extracted from the SWI volume.
   • Training utilizes a nested 3x5-fold cross-validation on 30 subjects (10 from an internal CAA cohort and 20 from a public dataset by Dou et al.), employing data augmentation (flips, rotations, intensity scaling/shift) and weighted sampling to address class imbalance.
   • This stage prioritizes high recall (sensitivity), casting a wide net to capture potential lesions.

3. Stage B: False-Positive Refinement:

   • A second, identical 3D ResNet classifier refines the predictions from Stage A.
   • Only patches with a Stage A probability ≥ 0.5 are forwarded to Stage B.
   • This stage is trained to discriminate true CMBs from the remaining false positives, prioritizing specificity.
   • The final output is a binary CMB segmentation NIfTI image and radiologist-ready QA visualizations.

Key Contributions

• Subject-Adaptive GMM Thresholding: A method to define intensity thresholds dynamically per subject, improving generalizability across different scanners and protocols without hardcoded cutoffs.
• Weakly Supervised Training: The pipeline is trained using only binary "yes/no" quality assurance labels for candidate patches, avoiding the need for labor-intensive dense voxel-wise segmentation masks.
• Two-Phase Cascade Architecture: A design that explicitly trades sensitivity for specificity, using Stage A for high recall and Stage B for high precision, achieving a balance suitable for clinical screening.
• Robustness to High Lesion Burdens: The system is validated on a CAA cohort with high microbleed counts, a scenario where previous models (e.g., SHIVA-CMB, MAGIC) have underperformed.

Results
The model was evaluated on a dataset of 11,424 candidates across 30 subjects.

• Performance Metrics: The cascade achieved an Area Under the Curve (AUC) of 0.9587, with a sensitivity of 0.712, specificity of 0.975, Positive Predictive Value (PPV) of 0.676, and an F1 score of 0.693.
• False Positive Reduction: The cascade reduced Stage A false positives by 76.8% (from 1,157 to 269) while retaining competitive sensitivity.
• Clinical Validation: Inference was performed on 141 SWI scans from a CAA clinical cohort. The algorithm detected a mean of 40.3 CMBs per scan. In a blinded comparison against the SHIVA-CMB model on 20 high-burden scans, a neurologist preferred the proposed algorithm's segmentations in 17 of 20 cases. The proposed method detected significantly more CMBs than SHIVA-CMB in high-burden cases (p < 0.05).
• Generalizability: The pipeline demonstrated robustness across 29 different SWI acquisition protocols.

Significance and Claims
The authors claim that this work provides a foundation for automated CMB detection in clinical settings, particularly for CAA patients who present with high lesion burdens. The significance lies in the ability to process heterogeneous clinical data without requiring dense manual annotations for training. By combining geometric filtering with a two-stage deep learning cascade, the system achieves high specificity (0.975), which is crucial for reducing the manual review burden on radiologists. The authors note that while the sensitivity (0.712) is moderate, the high specificity ensures the tool is applicable in clinical settings where false positives are costly. They suggest that the pipeline serves as a high-throughput quantification tool where missed lesions can be manually augmented, and the high specificity ensures strong clinical utility. The work also highlights the limitations of current public datasets and the necessity of internal cohorts for training models capable of handling the high burden of CAA.