AI-Assisted Brain Tumor MRI Reporting and Treatment-Planning Segmentation: A Retrospective Paired Workflow Evaluation.
Authors
Affiliations (13)
Affiliations (13)
- Institute of Biophotonics, National Yang Ming Chiao Tung University, Taipei 112, Taiwan.
- Department of Biomedical Science and Technology, National Sun Yat-sen University, Kaohsiung 804, Taiwan.
- Department of Radiation Oncology, Taichung Veterans General Hospital, Taichung 407, Taiwan.
- Department of Neurosurgery, Neurological Institute, Taichung Veterans General Hospital, Taichung 407, Taiwan.
- Department of Radiation Therapy and Oncology, Shin Kong Wu Ho-Su Memorial Hospital, Taipei 111, Taiwan.
- School of Medicine, Fu Jen Catholic University, New Taipei City 242, Taiwan.
- Division of Neuroradiology, Department of Radiology, Taichung Veterans General Hospital, Taichung 407, Taiwan.
- Department of Radiology, Taipei Veterans General Hospital, Taipei 112, Taiwan.
- School of Medicine, College of Medicine, National Yang Ming Chiao Tung University, Taipei 112, Taiwan.
- Department of Post-Baccalaureate Medicine, National Chung Hsing University, Taichung 402, Taiwan.
- Brain Research Center, National Yang Ming Chiao Tung University, Taipei 112, Taiwan.
- College Medical Device Innovation and Translation Center, National Yang Ming Chiao Tung University, Taipei 112, Taiwan.
- Center for Smart Health and Medicine, Taipei City Hospital, Taipei 112, Taiwan.
Abstract
<b>Background</b>: Brain tumor magnetic resonance imaging (MRI) reporting and tumor segmentation for treatment planning are time-consuming and variable. This retrospective fixed-sequence paired workflow study evaluates whether AI assistance is associated with changes in efficiency, consistency, and reproducibility. <b>Methods</b>: Thirty MRI cases (10 vestibular schwannomas, 10 meningiomas, 10 brain metastases) were assessed. Two neuroradiologists completed diagnostic reporting with and without AI assistance, and two physicians completed tumor delineation with and without AI-generated preliminary contours after a 3-week washout. <b>Results</b>: Reporting time decreased from 42.94 to 27.90 min for Reader A and from 101.04 to 80.47 min for Reader B, corresponding to median paired case-level reductions of 40.39% and 11.51%, respectively; only Reader A reached statistical significance. Sensitivity remained 97.73% and 100.00%, while precision was numerically higher after AI assistance (89.58% to 97.73% and 83.02% to 91.67%). Report-similarity metrics increased across ROUGE-L, BERTScore F1, and Sentence-BERT cosine similarity (all <i>p</i> < 0.001). Contouring time decreased from 54.63 to 4.93 min for Reader 1 and from 184.44 to 44.19 min for Reader 2, with median paired reductions of 100.00% and 87.11%. Dice coefficients were numerically higher after AI assistance (0.81 to 0.87 and 0.83 to 0.87). <b>Conclusions</b>: AI assistance was associated with shorter task-completion times, higher report-similarity metrics, and numerically higher contour-overlap measures. Prospective validation should determine whether these workflow efficiency gains translate into broader clinical benefit.