P10.25.A STANDARDIZATION AND AUTOMATIZATION OF MEASURING AND REPORTING BRAIN METASTASIS OVER TIME BY LEVERAGING ARTIFICIAL INTELLIGENCE Article Swipe
YOU?
·
· 2024
· Open Access
·
· DOI: https://doi.org/10.1093/neuonc/noae144.201
· OA: W4403510354
BACKGROUND Longitudinal assessment and reporting of brain metastases (BM) are prerequisites for a successful treatment regimen but exhibit inter-observer variability and are laborious. This study aimed to evaluate the workflow efficiency gains facilitated by a PACS-integrated artificial intelligence-enabled Lesion Tracking Tool (AI-LTT) for longitudinal BM monitoring, focusing on the accuracy in comparison to fully manual measurements and their inter-rater variability on post-Gamma Knife radiosurgery (GKR) patients. MATERIAL AND METHODS In this retrospective study, follow-up images of patients with BM who underwent GKR at our institution were examined on a research instance of our PACS (AI Accelerator, Visage Imaging, Inc.). Within manual workflow, two board-certified neuroradiologists displayed up to eight studies of each patient and measured orthogonal lesion diameters manually. In AI-LTT workflow introduced at RSNA2023, a custom hanging protocol automatically selected, displayed, and 3D registered T1 gadolinium-enhanced MR sequences in up to eight studies. A nnU-Net trained on the BraTS-METS2023 dataset (492 studies) and validated on 158 post-GKR studies automatically detected and 3D segmented the BMs from which the maximum diameters are extrapolated per RANO-BM criteria and presented in an organizational chart. This enhanced report can be imported to the free-text MRI report as a CSV file, containing the longitudinally aligned lesions, the related image series, orthogonal diameters, and the percentual change of each brain tumor compared to the prior study. The neuroradiologists revised the AI measurements as needed. We recorded the time and number of mouse clicks for both workflows from study open until completion of lesion measurement and analyzed the inter-observer variability of the reader`s manual measurements based on the BM evaluation of a third neuroradiologist. RESULTS Compared to manual measurements involving 50 studies with 352 displayed brain tumors of ten patients, AI-LTT demonstrated a significant reduction of mean time (678.5vs.366.0 seconds, P<0.01; 103.6vs.52.1 seconds per lesion, P<0.01) and mean clicks (201.5vs.62.9 clicks, P<0.01; 29.2vs.9.0 clicks per lesion, P<0.01) across two neuroradiologists. Mean absolute difference of the reader`s manual diameter measurements was 1.733±1.878 mm whereas intra-class and Spearman correlation coefficients were 0.922 and 0.881, respectively. The mean Dice coefficient of nnU-Net segmentations was 0.704±0.311 and sensitivity and F1-score of correctly identified lesions (≥5mm) were 0.816 and 0.855, respectively. CONCLUSION The AI-LTT allowed for a substantially faster workflow of measuring BMs while maintaining accuracy. The planned translation of AI-LTT into clinical practice can enable lesion-specific treatment response monitoring and standardize the measurement and reporting of BMs.
