Integrating advanced AI into the clinical 3D workflow

Abstract

Commercial 3D segmentation packages often lack state-of-the-art performance, requiring extensive manual intervention and thus hampering clinical efficiency. In response, we propose deploying advanced Deep Learning-based segmentation techniques, specifically nnU-Net, integrated into our open-source AI platform, SimpleMind. Our goal is to demonstrate the integration of this advanced AI into the clinical 3D workflow and assess its impact on reducing manual editing. 22 CT scans, requested for kidney volumetrics in transplant donor assessments, were randomly selected from our clinical 3D Lab’s inventory. Our advanced infrastructure, comprising HL-7 interfaces, databases, nodes, and GPU servers, enabled automated identification of the appropriate CT scans and series based on CPT codes and labeling logic. These scans were transferred from the clinical PACS to a secure research server running the SimpleMind AI platform. The segmentation results were returned to the clinical PACS as DICOM-seg objects for review and editing by the 3D image analyst on the Visage PACS 3D workstation. Annotation times per case were recorded for both manual and AI-assisted segmentation editing. The mean and standard deviation were calculated to evaluate the overall effort reduction and the impact of AI integration into the clinical workflow. The mean time for manual annotations was 11.5/11.0 ± 3.7 minutes, with a range of 7.0 to 19 minutes. For AI-assisted annotations, the mean/median time was 3.9/3.0 ± 1.0 minutes, ranging from 1.5 to 5.0 minutes. The percent reduction in annotation time, averaged 67.6%, with a maximum of 89.5% and a minimum of 37.5%. The reduction in annotation time was statistically significant (p < 0.05) by the Wilcoxon Signed Rank test.

Ähnliche Arbeiten