AI-driven target definition using CE-T1w and black blood sequence imaging in stereotactic radiosurgery for brain metastases.
Authors
Affiliations (4)
Affiliations (4)
- Radiotherapy Department, Isala, Zwolle, the Netherlands; Radiology and Nuclear Medicine Department, Isala, Zwolle, the Netherlands. Electronic address: [email protected].
- Radiology and Nuclear Medicine Department, Isala, Zwolle, the Netherlands.
- Medical Physics Department, Isala, Zwolle, the Netherlands.
- Radiotherapy Department, Isala, Zwolle, the Netherlands.
Abstract
Development and evaluation of a deep learning-based method for automatic detection and target delineation of brain metastases on contrast-enhanced T1-weighted (CE-T1w) and black blood (BB) MR scans for optimization of stereotactic radiosurgery planning. In this retrospective, single-center study, CE-T1w and BB CE-T1w scans with manual delineations from 206 consecutive stereotactic radiosurgery (SRS) patients were used to develop an nnUNet-based model. 162 patients were randomly allocated to the training set and 44 to the testing set. Detection was assessed using F1 score, sensitivity, and positive predictive value (PPV) across small (S; <0.1 cc), medium (M; ≥0.1 and < 1cc), and large (L; ≥1cc) lesions. Delineation was evaluated using Dice coefficient and Hausdorff distance. Multi-expert review by radiation oncologists and neuroradiologists served as reference. The AI model achieved an F1 score of 0.93 vs 0.96 by the experts. Sensitivity and PPV were 0.89 (AI) vs 0.97 (experts) and 0.98 (AI) vs 0.95 (experts). AI and clinicians showed similar sensitivities for M and L lesions (0.97-0.98 vs 1.0), while for small lesions, experts outperformed AI (0.95 vs 0.83). Overall, the experts missed 3.0% of the lesions. Dice coefficients were 0.76 (S), 0.90 (M), and 0.94 (L), averaging 0.83. The average Hausdorff distance was 1.4 mm. Our deep learning-based method shows promise for automated detection and delineation of brain metastases ≥ 0.1 cc on MRI in patients undergoing SRS. The AI model identified 3.0% of lesions missed by experts. Improving model performance, especially for small metastases, is key for clinical adoption.