Abstract 136: Reliability of an Automated, Machine‐Learning Algorithm for Volumetric Measurement of Patients with Chronic Subdural Hematoma

Background and PurposeHematoma volume in chronic subdural hematoma (CSDH) may predict neurological deterioration and need for surgical evacuation. Automated machine learning has shown promise in neuroimaging; however, the accuracy of automated subdural hematoma (SDH) volume measurement has not been compared to the Region‐of‐Interest (ROI) method.Materials and MethodsA total of 45 consecutive patients with CSDH from 2018 to 2019 admitted to our institution were selected. Exclusion criteria included: patients with bilateral SDH, slice size other than 512 times 512, non‐axial orientation, and post‐operative patients. The Region of Interest (ROI) method on a Horos Open‐Source Medical Image Viewer (Version 3.3.6) was utilized for CSDH volume measurement on CT scan by an experienced operator (NB, Figure 1). The Viz RECRUIT SDH (Viz.AI, Inc.) algorithm was used for automated, machine‐learning measurements. Inter‐observer reliability of volume measurements using the Bland‐Altman method was analyzed between the two methods.ResultsA total of 23 patients out of 45 (51%) were included in the analysis. Mean age was 77 years (range 48 ‐ 92 years). The mean hematoma volume was 76.6 cm3 (range 7.12 ‐ 160.22 cm3) for the ROI group and 105.42 cm3 (range 0.13 ‐ 207.41 cm3) for the Viz method. The Bland‐Altman coefficient of variation was 26.70% (range 1.73 ‐ 196.87 %), and all the data points were within the limits of acceptance (Figure 2).ConclusionThe inter‐observer analysis of CSDH hematoma volume using Horos Medical Image Viewer ROI and Viz RECRUIT SDH methods has good inter‐observer reliability. In contrast with the Viz RECRUIT SDH method, the ROI method is applicable to CTH with complex features including variable image resolution, bilaterality and post‐operative changes.