MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel
  1. Home
  2. Meeting Program
  3. All Sessions
  4. All Therapy Scientific Session
  5. A Deep-Learning Approach to Predict Planning Objectives in Cervical Cancer High-Dose-Rate Brachytherapy with a Tandem-And-Ovoid Applicator

A Deep-Learning Approach to Predict Planning Objectives in Cervical Cancer High-Dose-Rate Brachytherapy with a Tandem-And-Ovoid Applicator

Y Gonzalez1*, C Shen2 , P Klages3 , K Albuquerque4 , X Jia5 , (1) University of Texas Southwestern Medical Center, Dallas, TX, (2)University of Texas Southwestern Medical Center, Dallas, TX, (3) Memorial Sloan-Kettering Cancer Center, New York, NY, (4) University of Texas Southwestern Medical Center, Dallas, TX, (5) University of Texas Southwestern Medical Center, Dallas, TX,

Presentations

(Monday, 7/30/2018) 4:30 PM - 6:00 PM

Room: Davidson Ballroom A

Purpose: Treatment planning of high-dose-rate brachytherapy (HDRBT) is often performed under a high time pressure. It is highly desirable to predict treatment planning objectives based on patient-specific anatomy, which will effectively serve as guidance to treatment planning. Motivated by recent advances in deep learning area, this study develops a deep-learning approach to predict planning objectives in cervical cancer HDRBT with a tandem-and-ovoid (T/O) applicator.

Methods: A total of 145 cervical cancer cases previously treated at our institution under T/O HDRBT were selected. As the prescription dose varies among patients, all plans were first normalized to a reference prescription dose of 600 cGy. For each case, organ contours (bladder, rectum, sigmoid, and CTV) were extracted and distance map to CTV boundary was calculated. For each organ, we computed distance histogram of voxels inside the organ. Meanwhile, D2cc of each organ was extracted. We constructed a fully connected convolutional neural network (CNN) to map from distance histograms of all organs to D2cc of them. 132 cases were used for training with 10-fold cross validation. Each fold used a unique set of 13 validation cases that were not included in the training process. The remaining13 cases were set aside for testing.

Results: After training the network for 500 epochs, the trained network was able to predict D2cc with average difference of 3% on training data set averaged over the 10-folds. When applying the trained network to the testing data, difference was 8%. Individually, the bladder D2cc was predicted with 5% difference, the rectum D2cc within 10% and the sigmoid D2cc within 8%.

Conclusion: The developed deep-learning approach was able to predict the D2cc to each of the organs based on patient-specific distance histogram. Successful prediction of these variables is expected to facilitate treatment planning.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email

Additional Links