Encrypted login | home

Program Information

Rectal Toxicity Prediction in Cervical Cancer Radiotherapy Using Support Vector Machine and Deformable Accumulated Surface Dose

no image available
J Chen

J Chen1*, H Chen1 , Z Zhong2 , B Hrycushko3 , L Zhou1 , S Jiang3 , K Albuquerque3 , X Zhen1,3 , X Gu3 , (1) Southern Medical University, Guangzhou, Guangdong, (2) Wayne State University, Detroit, MI, (3) The University of Texas Southwestern Medical Center, Dallas, TX


TU-H-FS4-1 (Tuesday, August 1, 2017) 4:30 PM - 6:00 PM Room: Four Seasons 4

Purpose: This study aims to reveal dose-toxicity relationship via an accurate rectum surface dose accumulation method and a machine learning based prediction model for cervical cancer patients received combined high-dose-rate brachytherapy (BT) and external beam radiotherapy (EBRT) treatments.

Methods: The deformation vector fields that used to deformedly sum the rectum surface dose are obtained by registering the fractional rectum surface meshes to a reference domain by a recent developed local topology preserved non-rigid point matching algorithm (TOP-DIR). The accumulative equivalent 2-Gy three-dimensional rectum surface dose (3D-RSD) is flattened and mapped to a plane to obtain the two-dimensional rectum surface dose map (2D-RSD). Two types of dosimetric features, including 60 dose volume parameters and 600 dose geometric parameters are respectively extracted from the 3D-RSD and 2D-RSD. The Mann-Whitney test is performed to screen out the statistically significant features to further train a support vector machine model embedded with sequential forward search for features selection. The proposed model is validated on 42 cervical cancer patients (12 toxicity and 30 non-toxicity patients) with a leave-one-out method.

Results: The TOP-DIR achieved high accuracy in rectum surface matching with mean DC increases from 0.71 to 0.86, PE, VVD and HD decrease from 0.62, 1.5mm and 7.0mm to 0.26, 0.7mm and 4.0mm. The prediction accuracy, sensitivity, specificity and AUC of the proposed model are 80.2%, 55.4%, 90.0% and 0.83, respectively. Eleven dose geometric parameters are learned (with occurrence >60% in the leave-one-out validation) as predictors for toxicity prediction, but no dose volume parameters are selected, which suggests that the 2D dose geometric parameters are more meaningful for the rectum toxicity prediction than those 1D dose volume parameters.

Conclusion: The developed method is reliable and accurate in rectum toxicity prediction, which can serve as a potential tool for rectum complication prediction and treatment plan adaptation.

Funding Support, Disclosures, and Conflict of Interest: This work was supported by Varian Medical Systems Inc (#OTD-109235)

Contact Email: