Automatic longitudinal assessment of brain metastases improves detection of disease progression

Background: Characterization of multiple brain metastases (BMs) during treatment is limited by lack of available tools for individual volumetric lesion tracking. This study investigated role of a Picture Archiving and Communication System (PACS)-integrated artificial intelligence (AI) prototype lesion tracking tool (AI-LTT) for segmentation, enabling automatic assessment of longitudinal lesion changes compared to manual analysis by 2 neuroradiologists, focusing on inter-observer variability. Methods: Longitudinal studies of patients with BMs who underwent stereotactic radiotherapy were assessed. In manual workflow, 2 board-certified neuroradiologists displayed current and up to 7 prior studies of each patient and measured orthogonal lesion diameters (DMAX) manually. In AI-LTT-assisted workflow, a custom hanging protocol automatically selected, displayed, and 3D registered T1 gadolinium-enhanced MR sequences in up to 8 studies. A Scalable and Transferable U-Net (STU-Net) trained on BraTS METS 2023 BM dataset was used for segmentations from which DMAX were extrapolated. Neuroradiologists revised AI measurements as needed. Time and number of mouse clicks for both workflows were recorded. Results: A total of 40 patients with 158 MRIs were investigated. Median DSC of STU-Net segmentations was 0.771. Intra-class and Spearman correlation coefficients of manual diameter measurements were 0.959 and 0.961, respectively. Out of 353 true lesions, 59 were detected by only one (R1) and 15 by only the second neuroradiologist (R2). Among 74 lesions discrepant between readers, 14 proceeded to complete response, 4 partial response, 39 stable disease, but 12 developed progressive disease. Conclusions: AI-assisted workflow for BM analysis improves follow up of lesions and has potential to improve detection of lesions that progress over time of treatment.