MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Deep-Learning Based Prediction of Physicians Intention for High-Dose-Rate Brachytherapy with Tandem-And-Ovoids Applicator

Y Gonzalez*, C Shen, K Albuquerque, X Jia, The University of Texas Southwestern Medical Ctr, Dallas, TX

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: A wide range of plan evaluation metrics have been developed for radiotherapy treatment planning. However, none of them can truly reflect a physician’s intention during treatment planning, as it is hard to quantitatively represent physician’s mind. Driven by the mathematical foundation of deep learning that can use a deep neural network to represent functions accurately and flexibly, we developed a deep-learning framework that learns, and incorporates a physician’s intention to justify plan quality for cervical cancer high-dose-rate brachytherapy (HDRBT).


Methods: The system consisted of a dose prediction networks (DPN) and preference prediction network (PPN). DPN predicts EQD2 of OAR D2cc and CTV D90 from patient 3d anatomy. PPN reflects a physician’s intension by outputting the probability of a given plan being acceptable to the physician based on patient’s anatomy. Training of the networks was achieved in two steps. In the first step of individual training, DPN was trained to output clinical dose distributions, while PPN was trained to differentiate clinical plans from those generated by randomly perturbing dwell times of corresponding clinical plans. This step provided good initializations for the next step. In the second step of joint training, two network models were simultaneously finely tuned via an adversarial process. We collected approved treatment plans of 228 treatment fractions from 64 patients. Among them, 200 plans from 57 patients were employed as training and the remaining 28 plans from other 7 patients were saved for testing.


Results: DPN predicted EQD2 with an error of 8%±7.5%, 8.8%±5.9%, 7.2%±6.2%, and 6.9%±4.8% for bladder, rectum, sigmoid, and CTV, respectively. For PPN, the sensitivity, specificity, and AUC to predict physician’s decision to accept a plan were 0.90, 0.88, and 0.85.


Conclusion: We developed a novel deep-learning framework that predicts a physician’s intention to accept a plan for HDRBT treatment planning of cervical cancer.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email