3674 - Deep Learning Guided Direct Isodose Line Rendering for Stereotactic Brain Radiosurgery
Presenter(s)
L. Ma1, R. Liu2, A. Kalev3, A. Ma3, T. Zhang3, P. Dang1, S. Cheng2, and E. L. Chang4; 1USC Radiation Oncology, Keck School of Medicine, Los Angeles, CA, 2Department of Computer Science, University of Southern California, Los Angeles, CA, USA, Los Angeles, CA, 3Information Science Institute, University of Southern California, Los Angeles, CA, 4Department of Radiation Oncology, University of Southern California Keck School of Medicine, Los Angeles, CA
Purpose/Objective(s): Isodose lines are typically rendered and displayed only after a 3D dose computation in radiotherapy. In this study, direct rendering of isodose lines without 3D dose computation was developed via deep-learning (DL) guidance, and such a concept was specifically developed and tested on stereotactic brain radiosurgery.
Materials/Methods: An extended U-NET DL model was constructed first and trained over 500 brain radiosurgery cases to map target contours to a set (ranged from 1 to 5) of clinically approved plans of each case. Afterwards, DL-rendered isodose lines were filtered slice-by-slice to ensure their satisfaction of hardware-specific constraints for deliverability. To validate such an approach, DL-rendered isodose lines were compared with those of actual treatment delivery (ground truth) for a cohort of 100 brain radiosurgery cases of variable target sizes (1-2 cm in diameter), shapes and anatomical locations. The comparison was performed with dosimetric parameters such as the target volume coverage, the conformity index (CI) and the gradient index (GI).
Results: For the cases tested, isodose lines via DL guidance as described above were successfully rendered and displayed within minutes on associated MR image data. On average, the DL rendered isodose lines agreed excellently (<0.02 and <5%) with the ground truth (the second value in the bracket as follows), i.e., in the target volume coverage (mean 99.1% vs 99.0%), CI (mean 1.30 vs 1.25), and GI value (mean 2.5 vs 2.7). For some cases, the DL-rendered isodose lines exhibited better values and plan quality compared to the ground truth.
Conclusion: Direct rendering of isodose lines has been demonstrated for the first time for stereotactic brain radiosurgery. Further investigations with larger data sets and additional DL architectures are warranted toward end-to-end quality assurance and robust workflow integration of such an approach for clinical implementations.