Abstract WP227: Integration of Open Source Artificial Intelligence Based Intracranial Hemorrhage Volume Measurement and Prediction Tools into Clinical Workflows.

Background and Aims: Accurate hematoma volume measurements facilitate prognostication and translation of Intracranial Hemorrhage (ICrH) research into clinical practice. We built a deep learning network for automated segmentation/measurement of intracerebral (ICH), intraventricular (IVH) and extra-axial (EAH) components on non-contrast CT (NCCT) and integrated it into the clinical workflow of a statewide TeleStroke Service (TSS). The primary aim was to demonstrate improved model accuracy with prospective clinician adjustment of segmentations and retraining. A second aim was to develop a model for prediction of change in ICrH volume at 24h. Methods: A 3D convolutional neural network (SegResNet) was trained using the Medical Open Networks for Artificial Intelligence (MONAI) framework and 305 NCCT scans from a recently completed trial that included ICH patients within 6 hours of onset. Initial ICrH scans (120) from the TSS were used as an independent test set. The model was deployed with dicom connectivity to the TSS PACS, allowing inference segmentation and volume measurements to be assessed by clinicians assessing ICrH patients. A multimodal model including baseline NCCT, blood pressure, anticoagulant use, GCS and NIHSS scores was developed to predict ICrH expansion. Results: Initial median [95% CI] DICE scores for total ICrH were 0.82 [0.81,0.84]. ICH segmentation (0.83 [0.80,0.85] was more accurate than that for IVH (0.25 [0.14,0.38]) or EAH (0.23 [0.14,0.34]). Bland-Altman analysis indicated a mean bias of 4.3 ml (95% limits of agreement -18.6,15.9 ml). The model was used in 230 TSS patients over a 2-year period. Inference segmentations were assessed for accuracy, edited by clinicians and used to retrain the model. Median DICE scores improved to 0.88 [0.85,0.88], p