J Jiang*, S Alam, I Chen, P Zhang, A Rimner, J Deasy, H Veeraraghavan, Memorial Sloan Kettering Cancer Center, New York, NY

• 
  J Jiang

SU-IePD-TRACK 3-6 (Sunday, 7/25/2021) 12:30 PM - 1:00 PM [Eastern Time (GMT-4)]

Purpose: To train a AI auto-segmentation of lung tumors on weekly cone-beam (CBCT) using planning CT (pCT) and its clinical contour.

Methods: We developed a novel Patient-specific Anatomic Context and Shape prior information or PACS-aware 3D recurrent registration-segmentation deep network for segmenting lung tumors from CBCT. Using an end-to-end trained network, the anatomic context from higher contrast planning CT and shape (delineated contours) priors are incorporated as additional input channels into a 3D recurrent CBCT segmentation network through progressively deformed sequence of images and contours, produced by the recurrent registration network. Our approach learns to progressively deform images and handles large deformations between pCT and a weekly CBCT without requiring deformation vector field (DVF) for training. The segmentation network losses are optimized using CBCT expert segmentations. We trained our model using 98,000 pCT and CBCT image patches of size 192 x 192 x 48 extracted from 315 CBCT scans from 55 patients in a cross-validation setting. Independent testing was performed on 54 CBCT scans obtained from 10 patients. Performance comparisons was done against multiple state-of-the-art methods using Dice similarity coefficient (DSC), surface DSC and Hausdorff distance at 95th percentile (HD95) metrics.

Results: Our method produced significantly (p < 0.001) more accurate tumor segmentations (DSC of 0.84 ± 0.08, surface DSC of 0.98 ± 0.04, and Hausdorff distance at 95th percentile of 3.33±2.02mm) than both deformable image registration and deep learning voxelmorph methods on independent test set.

Conclusion: We introduced a patient-specific anatomic context and shape prior aware multi-modal recurrent registration-segmentation network for segmenting tumors on treatment CBCTs. Our approach significantly outperformed multiple methods and achieved promising accuracies on noisy and artifact-abundant cone-beam images.

Funding Support, Disclosures, and Conflict of Interest: Supported by the MSK Cancer Center support grant/core grantP30 CA008748. + NCI 5 R01 CA198121-04. Harini got fundings from Varian. Joseph Deasy is cofounder and share holder of Paige.AI. Rimner reports grants from Varian Medical Systems, grants and personal fees from AstraZeneca,Merck, Boehringer Ingelheim, grants from Pfizer

ePosters

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email