LAUSR

Background and Purpose: Intracerebral hemorrhage (ICH) expansion is an independent predictor of mortality and functional outcome with each milliliter of expansion increasing the chance of functional dependence by up to 7%. Unfortunately, detection of ICH expansion is often subjective, inaccurate, and may misguide treatment pathways. Artificial intelligence with convolutional neural networks (CNNs) represents a powerful new technology for image analysis and quantification. This study compares the accuracy, sensitivity, and specificity between a CNN optimized for ICH volume quantification and a traditional ABC/2 method. Materials and Methods: We performed a retrospective analysis of ICH patients who have had at least one follow-up non-contrast head CT (NCCT) within 24 hours. ICH expansion was defined as >33% volume of expansion, corresponding to a 10% increase in diameter. Each ICH was manually segmented, which served as ground truth measurements. Comparison of ICH expansion was made using (1) a traditional ABC/2 estimative approach and (2) a previously validated hybrid 3D/2D mask ROI-based CNN for ICH evaluation, which was trained previously on over 10,000 patients. Accuracy, sensitivity, and specificity of the CNN and ABC/2 approaches were then compared. Results: A total of 230 patients were included for a total of 460 NCCTs. The average ICH volume was 44.8 mL. The average ICH volume for the CNN was 45.3 mL (Pearson 0.99) and for ABC/2 was 60.4 mL (Pearson 0.81). Accuracy, sensitivity, and specificity for ICH expansion detection was 100%, 100%, and 100% for the CNN and 93.0%, 74.2%, and 96.0% for ABC/2. On visual inspection, cases of false positives by ABC/2 approaches tended to demonstrate eccentric expansion (Figure 1). Conclusions: A customized deep learning tool is highly accurate in the detection of ICH expansion. This may have important implications clinically for management and surveillance as well as in a clinical trial setting.