Improved automated cerebral microbleed (CMB) detection

Background: Cerebral Microbleeds (CMBs) are tracked in anti-amyloid drug safety monitoring. T2^*GRE or SWI MRI images are visually assessed for the presence of CMBs. Automated methods tuned for sensitivity often have unacceptably high false positive rates requiring extensive human cleanup. Method: Assume contemporaneous T2^*GRE and T1-weighted images. CMB detection (Figures 1,2) has three steps: 1) Segmentation outputs from the T1-weighted image are resampled into the T2^*GRE image voxel raster; 2) Normalized cross-correlation of the T2^*GRE with a CMB template identifies “candidate CMBs”; 3) a neural network using voxel patches centered on each candidate CMB rejects false positives. Patches presented to the network are 4D and include five contrasts drawn from the T2^*GRE, correlation, T1-weighted, WM probability and CSF probability images. The simple network architecture consists of convolutional blocks followed by dense blocks. The system is trained using images from ADNI, ARIC, MCSA, and the Mayo Clinic ADRC that have been previously visually graded for CMBs (Table in Figure 3). Importantly, we include participants with and without CMBs in our data reflecting the fact that the majority of people (even the elderly) have no CMBs. Data is split 70:10:20 by subject for training, validation, and testing. Various network configurations were evaluated. Result: High candidate detection sensitivity can be achieved at the expense of presenting hundreds to thousands of false positive candidates per input image to the rejection network. Our best performing network configuration to date has AUC 0.982 (95% bootstrap CI 0.971 to 0.990 based on the 10% validation subset). Choosing an operating point with 95% sensitivity in the validation set and admits, on average, 0.6 false positive CMB candidates per input T2^*GRE image. Sensitivity is hindered by a small number of obvious errors in the ground truth visual assessment. Validation loss tracks training loss with no indication of over-fitting. Example patches are shown in Figure 4. Conclusion: Excessive false positive detection requires significant human correction and effort, preventing automated CMB detection from being widely adopted. We have an automated detection system that is sensitive (95%) and provides excellent false positive rejection (<1 false positive CMB/input image) reducing needed human effort.