Toward Automated MR-Only Planning of the Brain: Organ-at-Risk Segmentation
Orasanu E, Glide-Hurst C, Stehle T, Buerger C, and Renisch S. Toward automated MR-only planning of the brain: Organ-at-risk segmentation. Med Phys 2017; 44(6):3029.
Purpose: Recent developments in synthetic CTs (synCTs) derived from MR data have enabled MR-only treatment planning of the brain to become more feasible. However, efficient and automated segmentation is still an unmet need. We aim at developing an automated model-based organ at risk (OAR) segmentation to support MR-only brain treatment planning. Methods: Twelve primary and metastatic brain cancer patients (5 post-surgical) underwent both CT-SIM and 1.0T MR-SIM (usual clinical protocol including T1w and T2w scans with resolutions of 0.9 × 0.9/1.25 mm3 and 0.7 9 0.7.2.5 mm3 respectively) within 1 week. Shape models were derived as triangle-based meshes for the skull, hemispheres, brainstem, optical nerves, globes, lenses and chiasm from the physician-delineated contours. Using a previously validated model-based segmentation framework, the brain model was trained for adaptation to both T1- and T2-weighted images separately. Once trained, the segmentation is fully automatic using a Generalized Hough Transform initialization. Segmentation performance was assessed by computing the distance between the clinical contours and the automatic segmentation result. Results: For the brainstem, we found a mean distance of 0.69 mm/0.63 mm between the clinical delineations and our segmentation on T1-w and T2-w images respectively (q95 1.83 mm T1/ 1.68 mm T2). For the optic model (lenses, optical nerves, lenses and chiasm) the mean distance was 0.66/0.73 mm (q95 1.79/1.94) on the T1-/T2-w images.For the optic system, the segmentation performs slightly better on T1-w than T2-w, which is also supported by a feature response analysis of the training and better resolution of the T1-w images. Conclusion: This study demonstrated the feasibility of developing automated OAR segmentation for MR-only brain treatment planning. Our results highlight the importance of acquiring data with high resolution to improve autosegmentation accuracy. Future work will include using a broader database, explore other sites such as head and neck. Overall, the work is promising for implementation in synthetic CT and MR-only workflows.