Automated detection of cerebral microbleeds on ex-vivo MRI scans of community-based older adults.
Authors
Affiliations (4)
Affiliations (4)
- Department of Computer Science, Illinois Institute of Technology, Chicago, IL, USA; Department of Biomedical Engineering, Illinois Institute of Technology, Chicago, IL, USA.
- Rush Alzheimer's Disease Center, Rush University Medical Center, Chicago, IL, USA.
- Department of Computer Science, Illinois Institute of Technology, Chicago, IL, USA.
- Department of Biomedical Engineering, Illinois Institute of Technology, Chicago, IL, USA; Rush Alzheimer's Disease Center, Rush University Medical Center, Chicago, IL, USA; Department of Diagnostic Radiology, Rush University Medical Center, Chicago, IL, USA. Electronic address: [email protected].
Abstract
Cerebral microbleeds (CMBs) are small hemosiderin deposits visible on T2*-weighted MRI that have been associated with cerebrovascular pathology, cognitive decline, and increased stroke risk. While CMBs have been studied extensively in living populations, their relationship to neuropathology assessed at autopsy remains incompletely understood. Large-scale MRI-pathology studies are needed to clarify these associations, but manual annotation of CMBs on ex-vivo MRI is time-consuming and labor-intensive, creating a critical bottleneck. Automated detection of CMBs on ex-vivo MRI in community-based older adults is particularly challenging due to low CMB prevalence, abundant mimics (e.g. air bubbles), and limited training data. We present the first comprehensive automated detection algorithm for CMBs on ex-vivo T2*-weighted MRI from community-based older adults. Our approach combines a novel multi-echo synthesis algorithm with self-supervised pretraining using fuzzy segmentation and confidence-aware learning to address data scarcity and class imbalance. The method successfully captures 90% of definite CMBs (unambiguous hypointensities clearly within brain tissue) and 83% of all CMBs (definite and possible CMBs combined) at 15 false positives per scan in a dataset of 287 community-based older adults. This represents a 46% improvement in average precision over the baseline approach using only real data and establishes a benchmark for this challenging detection problem. The proposed system enables partially automated annotation workflows that reduce manual review burden by 5 to 20-fold compared to feature-based approaches, making large-scale MRI and pathology studies feasible.