Click here to

Session: Multi-Disciplinary General ePoster Viewing [Return to Session]

Auto-Segmentation for Limited Field of View CBCT in Male Pelvic Region Using Deep Learning Method

H Hirashima1*, M Nakamura2, K Imanishi3, M Nakao4, T Mizowaki5, (1) Kyoto University, Graduate School of Medicine, Department of Radiation Oncology and Image-Applied Therapy, Kyoto, JP, (2) Kyoto University, Graduate School of Medicine, Department of Human Health Sciences, Kyoto, JP, (3) e-Growth Co., Ltd., Hyogo, JP, (4) Kyoto University, Graduate School of Informatics, Department of Systems Science, Kyoto, JP,(5) Kyoto University, Graduate School of Medicine, Department of Radiation Oncology and Image-Applied Therapy, Kyoto, JP

Presentations

PO-GePV-M-323 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: To develop an auto-segmentation model for cone-beam computed tomography (CBCT) with the limited field of view (FOV) in the male pelvic region using the deep learning method. Additionally, the effect of the difference of imaging device and FOV on the auto-segmentation accuracy using an external dataset was evaluated.

Methods: A total of 161 CBCT datasets from prostate cancer patients were included. Of them, 151 CBCT datasets were acquired by Vero4DRT (Hitachi), and 10 CBCT datasets were acquired by TrueBeam (Varian). The FOV was 20 cm and 26 cm for Vero4DRT and TrueBeam, respectively. Datasets were separated as follows; training dataset, 104 CBCTs: validation dataset, 27 CBCTs: test dataset, 30 CBCTs (20 CBCTs from Vero4DRT and 10 CBCTs from TrueBeam). Auto-segmentation model, U2-Net based on the combination of double U-net, was built. The ground truth region of interest (ROI) and predicted ROI were compared using the 3D dice similarity coefficient (DSC) and mean surface distance (MSD).

Results: Median (range: min-max) DSC was 0.96 (0.91–0.97), 0.84 (0.76–0.90), 0.87 (0.71–0.93), and 0.70 (0.36–0.83) on 20 CBCT datasets (Vero4DRT) and 0.94 (0.92–0.97), 0.85 (0.74–0.91), 0.84 (0.77-0.87), and 0.65 (0.32-0.88) on 10 CBCT datasets (TrueBeam) for bladder, prostate, rectum, and seminal vesicle, respectively. Median (range: min-max) MSD was 1.1 mm (0.7–2.3 mm), 2.2 mm (1.2–3.4 mm), 1.4 mm (0.7–5.4 mm), and 1.4 mm (1.0–3.0 mm) on 20 CBCT dataset (Vero4DRT) and 1.1 mm (0.8–1.4 mm), 1.8 mm (1.2-2.8 mm), 2.1 mm (1.7-3.4 mm), and 1.6 mm (0.8-3.1 mm) on 10 CBCT dataset (TrueBeam) for bladder, prostate, rectum, and seminal vesicle, respectively.

Conclusion: Auto-segmentation was successfully conducted by using the U2-Net model for limited FOV CBCT. The model developed in this study was robust to images acquired with different image capture devices.

Keywords

Cone-beam CT, Segmentation, Image Processing

Taxonomy

IM- Cone Beam CT: Segmentation

Contact Email

Share: