Deep Learning-Based Auto-Segmentation of Glioblastoma in Brain Cancer Radiotherapy
Presentations
PO-GePV-M-165 (Sunday, 7/25/2021) [Eastern Time (GMT-4)]
Purpose: Contouring is a potential source of uncertainties in radiotherapy treatment planning that could affect treatment outcomes. It is important to reduce contouring uncertainties in terms of inter-and intra-observer variation. In this study, we review the most recent researches and introduce an automatic brain tumor segmentation method. Our method is able to predict multiple brain tumor regions using the Brats 2019 dataset; We evaluate and report the dosimetric impact of automatic segmentation for treatment planning.
Methods: Two contouring methods are compared. We use manual contours as the gold standard against our deep-learning-based auto-contour. In each method, Gross Tumor Volume (GTV) was generated on 15 patients with glioblastoma cancer. Workflows were compared according to contour similarity, and dosimetric evaluation. Different treatment plans were compared using gamma index analyses were performed with 1%/1 mm, 3%/3 mm.
Results: The GTV segmentation network achieved a DSC of 0.88 ± 0.04 on the test set. The proposed network achieved a HD of 1.27 ± 0.39 mm for the GTV. The automatically generated contours met the dose coverage constraints using gamma analyses in all of the cases for GTV. Gamma index measurements for 1%/1 mm and 3%/3 mm were achieved at 97.92% and 99.375% respectively.
Conclusion: The proposed deep-learning-based auto-contour is fast, reliable, and achieved an acceptable agreement between the ground truth and predicted contours in relation to similarity and dosimetry. This method can provide a useful tool for automatic segmentation of the brain tumor and improve the reproducibility of contouring in the radiotherapy workflow.
Keywords
Absolute Dosimetry, Segmentation, Brain
Taxonomy
IM/TH- image Segmentation: Multi-modality segmentation
Contact Email