Abstract TP48: Automated Deep Learning-based Measurement of DWI Lesion Volume in Acute Stroke Using Convolutional Neural Networks

Abstract

Background: In acute ischemic stroke (AIS), therapeutic decisions are increasingly being based upon the volume of likely-unsalvageable brain tissue, which is often estimated using DWI. Deep learning algorithms, e.g. convolutional neural networks (CNN), have been employed for chronic stroke lesion segmentation. Here we investigate the applicability of CNN for DWI lesion measurement in acute stroke. Methods: 50 AIS patients underwent DWI < 12h from last known well. Apparent diffusion coefficient maps, T2WI, and DWI were used as covariates in a 2D CNN (5-fold cross validation). Including convolutional, inception and fully connected dense layers, a CNN of 15 layers was trained using manually outlined DWI lesions. To avoid overfitting, statistical dropout, L1- and L2-regularization and batch-normalization were used. Automatically segmented lesion volumes (ALV) using a 50% risk threshold were compared to the manual lesion volumes (MLV) using Dice similarity index (DSI, a measure of overlap) and Spearman’s correlation coefficient. Subset analysis was performed evaluating results between small (<10 ml) and large lesions (Wilcoxon rank sums). Results: The figure shows examples of CNN segmentation. The median [IQR] measured lesion volume and DSI were 25 [13-46] mL and 66% [35-75%], respectively. The correlation of MLV with ALV was 86% (P<0.001). 21 subjects (42%) had lesion volumes less than 10 ml. DSI for small lesions (28% [14-46%]) was significantly lower (P<0.001) than large lesions (73% [67-79]%). Correlation of ALV with MLV for small lesions compared to large lesions were 31 and 84 respectively and differed significantly (P=0.001). Discussion: Automatic DWI lesion segmentation for large lesions is feasible using CNN. CNN tended to overestimate the volumes of small lesions. Prior methods have used a priori heuristics or morphometric operations to remove artifacts. CNN methods show promise for “learning” to discriminate artifacts from real lesions.