Radiation and Cancer Physics

SS 18 - Physics 4 - Imaging for Response Assessment I

142 - Feasibility and Consistency of On-Treatment Cone Beam CT (CBCT) Feature Extraction Using an Automated Deformable Registration Radiomics Pipeline

Tuesday, October 23
8:25 AM - 8:35 AM
Location: Lila Cockrell Theatre

Feasibility and Consistency of On-Treatment Cone Beam CT (CBCT) Feature Extraction Using an Automated Deformable Registration Radiomics Pipeline
P. Lang1, T. Patel2, U. Bernchou3, C. Brink4, D. J. Moseley5, N. Becker5, T. Tadic6, and A. J. Hope7; 1Princess Margaret Cancer Centre, Toronto, ON, Canada, 2Princess Margaret Cancer Centre, Toronto, PE, Canada, 3Odense University Hospital, Odense, Denmark, 4Laboratory of Radiation Physics, Odense University Hospital, Odense, Denmark, 5University of Toronto, Toronto, ON, Canada, 6Princess Margaret Hospital, Toronto, ON, Canada, 7Department of Radiation Oncology, Princess Margaret Cancer Centre-University of Toronto, Toronto, ON, Canada

Purpose/Objective(s): To evaluate a novel automated pipeline (AP) with an alternate deformable image registration algorithm compared to semi-automated deformable registration (SADR) to derive CBCT imaging markers (CBCTM) previously correlated to radiation pneumonitis (RP).

Materials/Methods: NSCLC patients treated between 2011-2014 with curative radiation with ≥ 12 months follow-up without tumor recurrence prior to development of toxicity were included. RP (CTCAE G2+) was obtained from a prospectively collected database. Treatment plan and CBCT registrations for all treatment fractions was attempted via: (SADR) an intensity-based algorithm using the Elastix toolbox and custom pre-processing, or (AP) an intensity-based algorithm in RayStation with the CBCTM extracted in a custom image processing module. The SADR dataset was the training dataset, and extra patients extracted via AP were the validation dataset. Patient characteristics in the training/validation and correlations between CBCTM values (SADR and AP) were compared. Associations between RP, CBCTM at fractions 10 and 20 (CBCTM10, CBCTM20), average CBCTM of 5 consecutive fractions around fractions 10 and 20 (CBCTM10avg, CBCTM20avg), and dosimetric factors were evaluated using logistic regression. Model performance was evaluated using area under the receiver operating characteristic curve (AUC) analysis. Models derived from the training dataset were evaluated independently in the testing dataset.

Results: 129 patients were included, with 36 patients (27.9%) having RP. 83/129 (64.3%) were successfully registered using both SADR and AP (training set). Failures were due to inconsistent data formats, convergence failure, and grossly incorrect registrations. The remaining 46 patients were successfully processed using AP (validation set). Patient characteristics including the rates of pneumonitis (26.5% vs. 31.1%) in the training and validation datasets were similar. In the training dataset: CBCTM values extracted using SADR and AP were highly correlated for both F10 and F20 (rho =0.81, 0.92, p <0.005); using either SADR or AP, mean lung dose (MLD) and CBCTM20 were associated with RP on multivariable analysis (p<0.05); CBCTM10avg and CBCTM20avg were both correlated with RP (p<0.05) on univariable and multivariable analysis; and the best performing model (CBCTM20+MLD) had an AUC of 0.679. In the validation dataset, the CBCTM20+MLD model had an AUC of 0.678.

Conclusion: A previously validated CBCT marker of lung toxicity can be automatically derived using a different deformable image registration method, and remains correlated with RP, in addition to high correlation of marker values between the two methods. The CBCTM20 + MLD model was predictive of RP in an independent validation dataset.

Author Disclosure: P. Lang: None. T. Patel: None. U. Bernchou: None. D.J. Moseley: Employee; University Health Network. Royalty; Modus Medical Devices. N. Becker: None. T. Tadic: Royalty; Modus Medical Devices Inc. Patent/License Fees/Copyright; Modus Medical Devices Inc, Alberta Health Services, University Health Network. A.J. Hope: Travel Expenses; Elekta, Inc.

Pencilla Lang

Disclosure:
No relationships to disclose.

Presentation(s):

Send Email for Pencilla Lang


Assets

142 - Feasibility and Consistency of On-Treatment Cone Beam CT (CBCT) Feature Extraction Using an Automated Deformable Registration Radiomics Pipeline



Attendees who have favorited this

Please enter your access key

The asset you are trying to access is locked. Please enter your access key to unlock.

Send Email for Feasibility and Consistency of On-Treatment Cone Beam CT (CBCT) Feature Extraction Using an Automated Deformable Registration Radiomics Pipeline