Clinical Validation of Segmentation-Based Detection of Glioma Progression

Purpose To evaluate whether an AI-based method could be used routinely as part of patient care to assist in detecting non-enhancing glioma progression. Materials and Methods A 3D U-Net trained (n=481) and validated (n=121) to segment post-surgical lower grade gliomas was used to measure tumor volumes over time and assess progression in a clinical test set. Eight prospective and eight retrospective patients (total 72 exams) who were suspected of progression during their routine outpatient imaging were clinically assessed. Gold standards for progression were derived from clinical reports a posteriori using visual read, and radiologists were blinded to the AI decision at time of reporting. Results Progression assessments were presented to radiologists via an easy-to-use, interactive, and interpretable environment in under 10 minutes. Combining prospective and retrospective cases, a final sensitivity of 0.72 and specificity of 0.75 was achieved at progression detection. Conclusions Automated detection of glioma progression would provide valuable decision support for routine use.