IMPROVED CEREBRAL MICROBLEED DETECTION IN MAGNETIC RESONANCE IMAGES USING A MULTI-SCALE 3D CONVOLUTIONAL NEURAL NETWORK (ID 1074)
Abstract
Background And Aims
Cerebral microbleeds are an important neuroimaging biomarker for haemorrhage risk. Their utility in clinical practice is limited as manual detection from magnetic resonance images can be time-consuming and error-prone. Despite recent advances, automated detection remains a challenging task due to their small size, widespread distribution, and frequency of mimics such as veins and calcifications. This study aims to address these obstacles by developing and validating a fully automated approach utilising information at multiple scales to better detect cerebral microbleeds.
Methods
An efficient two-stage approach was developed: first, using a traditional three-dimensional radial symmetry transform stage to detect candidates; and second, using a novel three-dimensional convolutional neural network (3D-CNN) stage to reduce false positives. Compared with previous methods, our 3D-CNN incorporates larger contextual information by employing a dual pathway residual architecture that processes candidates at multiple scales simultaneously. The proposed approach was trained and evaluated using a recently released public dataset consisting of susceptibility-weighted imaging data for 780 microbleeds from 72 subjects with high and 107 subjects with low in-plane resolution.
Results
An overall 5-fold cross-validated sensitivity of 87.7% and 72.2% were achieved with an average of 1.33 and 1.79 false positives per subject for the high and low in-plane resolution data respectively.
Conclusions
The developed approach outperformed existing methods on the public dataset with comparable sensitivity but fewer false positives. It is hoped that this work will help reduce the burden of measuring microbleeds and facilitate better risk scores.
Trial Registration Number
Not applicable