MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel
  1. Home
  2. Meeting Program
  3. All Sessions
  4. All Joint Imaging-Therapy SNAP Oral
  5. A Deep Learning Approach of the Image Based Acquisition of the Position and the Shape of a Lung Tumor Using X-Ray Color Fluoroscopic Images

A Deep Learning Approach of the Image Based Acquisition of the Position and the Shape of a Lung Tumor Using X-Ray Color Fluoroscopic Images

Y Kajihara1*, T Shiinoki2 , F Fujii1 , (1) Yamaguchi University, Ube, Yamaguchi, (2) Yamaguchi University, Ube,Yamaguchi

Presentations

(Sunday, 7/29/2018) 4:00 PM - 4:55 PM

Room: Karl Dean Ballroom C

Purpose: Eyeing on the future implementation of the markerless real-time tumor identification using stereo x-ray fluoroscopic images, we tackle the problem of detecting tumor. The purpose of this study is to construct the markerless tumor tracking system to detect the position and the shape of a lung tumor using the deep neural network.

Methods: A patient who underwent respiratory-gated radiotherapy using SyncTraX were enrolled in this study. Seven fields were created for treatment plan. For each field, color fluoroscopic images were acquired at 30 fps for 10 s and total 4200 images data were acquired during the treatment (images x fields x X-ray tube). Color fluoroscopic images acquired during treatment simulation were classified into a similar groups and were randomly selected 100 images. Two hundred images for training were created based on selected 100 images. Twenty images for test were randomly selected from each group. The U-Net model was used to create the training model and that was verified using testing images. For RGB components, one component with clear lung tumor was used in this study. For every group, the dice coefficient was calculated to verify the constructed system for test images. The differences between the centroids of supervised images and those of the binarized output image were calculated as positional errors.

Results: For all groups, the mean ± standard deviation (S.D.) of dice coefficient in test images was 0.89 ± 0.05. Detection of tumor shape could be achieved with high accuracy. However, mean ± S.D. of the centroid positional errors were 2.50 ± 1.50, exceeding 1 mm.

Conclusion: We have constructed the markerless tumor tracking system to detect the position and the shape of a lung tumor using the deep neural network. It is needed to investigate the tracking accuracy for many clinical cases.

Keywords

Image-guided Therapy, Segmentation, Fluoroscopy

Taxonomy

TH- External beam- photons: Motion management (intrafraction)

Contact Email

Additional Links