Program Information
Image Segmentation for Tumor Tracking by Deep Learning with Robustness for Obstacle Object Feature
T Terunuma*, T Sakae , University of Tsukuba, Tsukuba, Japan
Presentations
TU-RPM-GePD-JT-6 (Tuesday, August 1, 2017) 3:45 PM - 4:15 PM Room: Joint Imaging-Therapy ePoster Theater
Purpose: To evaluate the validity of newly devised markerless tumor-tracking method based on Deep Learning (DL) with recognition controlling of importance-feature (tumor) and unimportance-feature (bone) for tracking.
Methods: A supervised DL was adapted to make the different recognition that tumor-feature is important for tracking and bone-feature is unimportant for tracking. The Input data-set for DL was a massive pair of training images and supervised images. The training images were randomly overlapped a bone DRR on a soft tissue DRR which included a projected tumor feature. The supervised images were the binary images which indicated the segmented tumor-shape. Using the created classifier by DL, tumor-shapes in fluoroscopy were segmented.
Results: Calculation time were 90 min in learning and 25 msec in segmentation. The similarity between segmented tumor-shape and true were about 0.95 (Jaccard Index) in geometric model test and in virtual fluoroscopy test. The correlation between tracked tumor trajectory and true were 0.99. Error of these model tests were about 1 mm. In clinical fluoroscopy test, error of tracking was preliminary estimated about 1 mm.
Conclusion: This work indicates the potential for real-time tumor tracking and segmentation.
Contact Email: