Click here to

Session: AI-Based Auto-Segmentation and Auto-Contouring - I [Return to Session]

Simultaneous Segmentation of Target and Organ at Risk in Thoracic Kilovoltage Images

M Mueller1,2*, A Mylonas1,2, P Keall1, J Booth3, D Nguyen1,2,3, (1) ACRF Image X Institute, Sydney, NSW, AU, (2) University of Technology Sydney, Ultimo, NSW, AU (3) Royal North Shore Hospital, Sydney, NSW, AU


TU-E-TRACK 6-6 (Tuesday, 7/27/2021) 3:30 PM - 4:30 PM [Eastern Time (GMT-4)]

Purpose: The ability to simultaneously localize the treatment target and local organs-at-risks (OARs) can improve SABR treatment efficacy and reduce side effects by ensuring the target dose and avoiding irradiating the OARs. We present a fiducial-less Multi-Target-Segmentation (MTS) approach that utilizes the standard-equipped kV-imager of a conventional linac.

Methods: MTS used a conditional generative adversarial network to simultaneously segment the tumor and the heart in kV-images. MTS was trained on the DRRs of the planning CT for each patient, such that it can be deployed for intra-fractional target tracking. A patient-specific training dataset consisted of 36000 DRRs, created by forward-projecting each of ten planning 4D-CT phases and segmented target volumes equidistantly over a full imaging rotation. The labels of multiple targets were stacked as separate image channels. Data was augmented during training by randomly cropping the field-of-view within ±3mm. MTS was trained individually for three lung SABR patients and tested on kV-images of three treatment fractions per patient. The ground truth target positions were determined through manual alignment of the forward-projected target volumes onto the kV-projections. The geometric accuracy was determined by comparing the centroid of the segmented targets on each image to the ground truths. Additionally, the dice-similarity-coefficient (DSC) was calculated to evaluate MTS.

Results: MTS had a mean tumor localization accuracy of -0.4±2.6mm and 0.7±1.7mm in the lateral/ anterior-posterior and superior-inferior directions of the images, respectively. The mean heart localization accuracy was 1.9±1.6mm and 1.4±1.1mm in the lateral/anterior-posterior and superior-inferior directions of the images, respectively. The mean DSC was 0.72±0.12 for the tumor and 0.88±0.04 for the heart, respectively.

Conclusion: Simultaneous target and OAR localization without fiducials is feasible and can achieve clinically relevant accuracy for clinical deployment. MTS may widen the therapeutic index of SBRT by minimizing thoracic toxicity for targets in close proximity of critical OARs.



    Segmentation, Image-guided Therapy


    IM/TH- image Segmentation: X-ray

    Contact Email