Clinical Implementation of the Pancreatic Cancer Fast CT-CBCT Registration Network Model
E LoCastro1*, J Hong2, Y Hu3, X Han4, A Apte5, G Mageras6, (1) Memorial Sloan Kettering Cancer Center, New York, New York, (2) MSKCC, New York, NY, (3) Memorial Sloan Kettering Cancer Center, New York, NY, (4) Unc Chapel Hill, ,,(5) Memorial Sloan-Kettering Cancer Center, Maywood, NJ, (6) Memorial Sloan-Kettering Cancer Center, New York, NY
Presentations
PO-GePV-M-34 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: We previously reported on a deep-learning (DL) deformable CT-CBCT registration model for segmentation of organs-at-risk (OARs) in breath-hold CBCT scans for pancreatic cancer radiotherapy. Here we report on the development of automated pre- and post-processing of the images to enable clinical evaluation of the DL model.
Methods: We automated three manually performed steps in the original algorithm: (1) determination of lower (L) and upper (U) bounds in CBCT voxel intensities such that intensities within the bounds are linearly remapped to a standardized interval [-1,1]; (2) determination of parameters to identify gas pockets for filling; and (3) rigid registration of planning CT to CBCT. In step (1), linear correlations were determined between mean and standard-deviation (SD) voxel intensities within a volume-of-interest and manually determined window-level (U+L)/2 and range U-L in a 40-patient test cohort (80 CBCT scans). In step (2), voxel intensity threshold (T) and multiplier (M) values for determining gas pockets were accomplished via figure-of-merit (FOM) calculation, which examined gas pocket volume, mean and standard deviation voxel intensities. The algorithm computes gas pockets for different combinations of (T,M) and uses FOM to select the parameter set from the best matched case in the test cohort. A pipeline was built to integrate the preprocessing and DL registration software by utilizing an in-house system (AWARE), which retrieves images, segmentations and initial CT-CBCT alignment parameters at treatment time (step 3). The pipeline passed data to a high-performance cluster for DL processing, followed by retrieval of deformed segmentation for review.
Results: Visual inspection of CBCT scans show organ visibility is similar between automatic and manually determined lower/upper intensity bounds. Automated vs manual gas volumes and OAR Dice coefficients show good agreement.
Conclusion: By developing sufficiently robust automated methods to determine parameters, we have enabled DL-based CT-CBCT registration model in pancreas without manual intervention.
Keywords
Treatment Planning, Nonlinear Image Warping, Cone-beam CT
Taxonomy
Contact Email