Using 3D multi-scale convolutional neural networks to automate the segmentation and quantification of acute ischemic lesion

The accurate segmentation of acute ischemic lesions in diffusion-weighted images (DWI) is crucial for treatment decisions related to ischemic stroke. Manual delineation is time-consuming and subject to interobserver variability. Deep learning has shown promise in infarct segmentation; however, existing methods are limited by the heterogeneity of ischemic lesions and DWI acquisition protocols. This study developed a 3D multi-scale convolutional neural network (CNN) to automate the process of infarct segmentation and volume quantification in DWI. This study retrospectively collected DWI data from 184 ischemic stroke patients (218 scans) at Taipei Veterans General Hospital (TPEVGH) for model development and validation. Brain parenchyma extraction was performed using a 2D mask region-based convolutional neural network (Mask R-CNN), while infarct segmentation was performed using a 3D multi-scale CNN (DeepMedic). Model performance was assessed using five-fold cross-validation of the validation set and three external test sets obtained from Landseed Hospital (LSH, n = 66 scans), Jen-Ai Hospital (JAH, n = 78 scans), and Taoyuan General Hospital (TYGH, n = 59 scans). Segmentation accuracy was quantified in terms of Dice coefficient, precision, and recall. When applied to the validation set for brain parenchyma extraction, model performance was as follows: mean Dice coefficient (0.974 ± 0.001), precision (0.987 ± 0.001), and recall (0.962 ± 0.001). When applied to the validation set for brain infarct segmentation, model performance was as follows: mean Dice coefficient (0.746 ± 0.012), precision (0.767 ± 0.014), and recall (0.768 ± 0.012). When applied to external datasets, the proposed model yielded Dice coefficients of 0.729 ± 0.148 (LSH), 0.707 ± 0.171 (JAH), and 0.684 ± 0.136 (TYGH). These findings demonstrate the robust generalizability of the model across institutions. Interestingly, segmentation accuracy was significantly higher when applied to MRI obtained at 4–7 days after stroke onset, compared to the day of onset. The proposed 3D multi-scale CNN demonstrated excellent accuracy and generalizability in infarct segmentation and volume quantification. The proposed automated segmentation scheme could significantly accelerate stroke imaging analysis, reduce interobserver variability, and enhance clinical decision-making in the management of acute ischemic stroke. Not applicable.